FREEWAY TRAVEL TIME ESTIMATION USING LIMITED LOOP DATA. A Thesis. Presented to. The Graduate Faculty of The University of Akron

Transcription

1 FREEWAY TRAVEL TIME ESTIMATION USING LIMITED LOOP DATA A Thesis Presented to The Graduate Faculty of The University of Akron In Partial Fulfillment of the Requirements for the Degree Master of Science Silin Ding May, 2008

2 FREEWAY TRAVEL TIME ESTIMATION USING LIMITED LOOP DATA Silin Ding Thesis Approved: Accepted: Advisor Dr. Ping Yi Department Chair Dr. Wieslaw K. Binienda Committee Member Dr. Ala R. Abbas Dean of the College Dr. George K. Haritos Committee Member Dr. William H. Schneider IV Dean of the Graduate School Dr. George R. Newkome Date ii

3 ABSTRACT Providing drivers with real-time, good-quality traveling information is becoming increasingly important as congestion increases in cities across the United States. Studies have shown as congestion increases, travel time reliability decreases. Travelers would like to have information about certain traffic conditions as particularly detours causing time delays, delays because of road constructions, and delays due to accidents. Since congestion is treated as a major factor influencing travel decisions, some metropolitan areas are providing travel time information to motorists through dynamic message signs (DMS), 5 programs, the Internet, highway advisory radio, and other sources. Traffic conditions are affected by current events and established travel patterns. Today, travel time data can be gathered from microwave radar, automatic vehicle tag matching, video detection, license plate matching, and most commonly, inductive loops. Loop detectors are placed in individual lanes to provide volume, occupancy and local speed information. Although closely spaced loop detectors are helpful to system operation, they are expensive to install and to maintain. With the proliferation of cell phone usage, loop detector data is no longer critical to incident detection. The effectiveness of using loop detector data to reliably estimate travel time has yet to be proved. In recent years, researchers discussed the pros and cons of detector spacing. This discussion is necessary and timely because of the widespread use of the loop detection system today. The focal iv

4 point of the discussion is to determine the appropriate detector spacing needed for various applications while maintaining the same level of data quality for all users. This thesis studied different freeway travel time estimation methods and explored the impact of loop detector spacing on travel time estimation. The analysis was performed on a sixteen-mile stretch of I-75 in Cincinnati, Ohio and used both simulation and field tests to evaluate the results. First, the commonly used midpoint method for travel time estimation was examined under various traffic and roadway conditions. Starting with the existing /3 mile spacing, spacing was increased by using fewer detectors to obtain data for analysis. Then, enhancements were introduced over the midpoint method using different data processing methods reported by other researchers to improve its performance. Preliminary results showed that by using the midpoint method, different detector spacings result in different levels of accuracy and generally the estimation error increases with the detector spacing. Moreover, with increasing traffic congestion, the travel time errors from the existing methods increased significantly. After a congestion based error correction term is introduced, the improved midpoint method is able to make more accurate travel time estimates at larger spacings under work zone and incident conditions. The work was also tested against field data collected through probe vehicles. Based on field data, the estimated travel times from the improved method matches closely with those measured by the floating cars; the differences between the travel time are within 0%. Results from this study showed that a larger detector spacing than the commonly used /3 mile does not worsen the estimation results. Overall, the one-mile spacing scheme has outperformed the other tested alternatives in the testbed area. iv

5 This thesis also studied the reliability of the probe vehicle technique. License Plate Matching Survey was conducted to carry out the analysis. The results showed that the accuracy of probe vehicle travel time is affected by the standard deviation of travel time and different analysis periods. Minimum sample size was examined as the last part of the thesis. v

6 ACKNOWLEDGEMENTS First and foremost, I would like to express my sincere appreciation to my advisor, Professor Ping Yi, for his valuable guidance, continuous support and encouragement throughout this study. The helpful advices and collaboration from Halle Jones Capers who was the director of Women in Engineering in the University of Akron are gratefully acknowledged. The useful data information from Andrew Fluegemann at ARTIMIS center in Cincinnati, OH is really appreciated. Acknowledgements are also extended to my committee members, Dr. Ala R. Abbas, Dr. William H. Schneider, for reviewing my work and helpful recommendations. Special thanks are given to my fellow graduate students, Chun Shao, Jialei Mao, and Cong Feng for their useful discussions related to the topic of travel time estimation and their support in data collection. The sincere friendship and support from them always gave me energy and impetus to finish this dissertation. My deepest gratitude goes to my family who provide love and support more than I could ever expect. vi

7 TABLE OF CONTENTS LIST OF TABLES... ix LIST OF FIGURES... x CHAPTER I. INTRODUCTION Statement of the Problem Research Objectives and Methodology Organization of the Thesis... 6 II. LITERATURE REVIEW Introduction Inductive Loop Detector Data (ILDD) Collection Technologies Freeway Travel Time Estimation (TTE) From ILDD Detector Spacing Issue and Freeway Travel Time Estimation Concluding Remarks... 4 III. FREEWAY TRAVEL TIME ESTIMATION AND COMPUTER MODELING Introduction Spot speed methods Average Speed Method Minimum Speed Method Distance/Volume Weighted Average Method... 7 vii

8 3...4 Midpoint Method Traffic Flow Theory Method Computer Modeling and Implementation History and Background of Traffic Simulation Introduction of VISSIM (Urban Traffic Simulator) Methodology and Data Analysis Concluding Remarks IV. FREEWAY TRAVEL TIME STUDY WITH FIELD DATA General Description of Field Data Collection Data Collection Data Processing Evaluation of Freeway Travel Time Estimation with Field Data Reliability of Probe Vehicle Method Introduction of Sample Size Requirements and License Plate Matching Method Data Source of License Plate Matching Data Collection Data Analysis Concluding Remarks V. CONCLUSION BIBLIOGRAPHY APPENDIX viii

9 LIST OF TABLES Table Page Table 3. Calibration Results on Different Types of Roadway Sections Table 3.2 One-Sample t-test P-Values, 95% Upper Confidence Bound for Midpoint Method (α=0.05) Table 3.3 Comparison of One-Sample t-test P-Values, 95% Upper Confidence Bound of Percentage Error (%) for Improved Method (α=0.05) Table 4. Summary of Data Collection Parameters Table 4.2 Standard Deviation and Coefficient of Variance of Travel Time Table 4.3 P-value of two-way ANOVA table (95% confidence interval) Table 4.4 Sample Sizes for 5-minute, 0-minute and 5-minute Analysis Periods Table 4.5 Sample Sizes for 5-minute, 0-minute and 5-minute Analysis Periods Based on Repetition Factor ix

10 LIST OF FIGURES Figure Page Figure. Typical Configuration of Loop Detectors on Freeways... 3 Figure 3. Illustration of Spot Speed Method... 7 Figure 3.2 Illustration of Continuity Equation Figure 3.3 A Freeway Links of I-75, I-7 and I-275 in Cincinnati Figure 3.3 B VISSIM Simulation of Entire I-75 Network Figure 3.4 Percentage Errors at Different Traffic Volume Levels... 3 Figure 3.5 Absolute Mean Error from Midpoint Algorithm Figure 3.6 Summary of Percentage Errors Figure 3.7 Absolute Mean Error of Different Detector Spacing under Different Traffic Volume Figure 4. GPS Equipment Figure 4.2 Polled GPS Data Figure 4.3 Speed comparison between GPS data and loop detector data Figure 4.4 Travel time estimation errors with floating car data Figure 4.5 Spreadsheet form of License Plate Numbers... 5 Figure 4.6 Matched Result of License Plate Figure 4.7 Percentage Errors under Different Number of Probe Vehicles for 5-Minute Analysis Period x

11 Figure 4.7 Percentage Errors under Different Number of Probe Vehicles for 5-Minute Analysis Period (Continued) Figure 4.8 Percentage Errors under Different Number of Probe Vehicles for 0-Minute Analysis Period Figure 4.9 Percentage Errors under Different Number of Probe Vehicles for 5-Minute Analysis Period xi

12 CHAPTER I INTRODUCTION Travel time is a fundamental measure in transportation engineering that can be provided to a wide variety of audience, including engineers, planners, administrators, and motorists through dynamic message signs (DMS), 5 programs, websites, highway advisory radio and other sources. As a measure of performance and decision-making variable, travel time is useful in many aspects of transportation planning, modeling, and decision-making applications. These applications include traffic and performance monitoring, congestion management, travel demand modeling and forecasting, traffic simulation, and traffic operations strategies. Travel time information is becoming increasingly important for a variety of real-time transportation applications. These realtime applications include Advanced Traveler Information Systems (ATIS), Route Guidance Systems (RGS), etc., which are parts of the Intelligent Transportation Systems (ITS). Providing travelers with real-time, high-quality travel time is becoming increasingly important as congestion continues to grow in cities across the United States. Travel time is affected by the current traffic condition and the established travel patterns. Today, real time data can be gathered from microwave radar, automatic vehicle tag matching, video detection, cell phone matching, license plate matching, probe vehicles, and most commonly, inductive loops. Loop detectors can be placed in each lane 2

13 to provide volume, occupancy and local speed information. Generating travel time estimates from fixed loop detectors is the most common method in the U.S. According to the Intelligent Transportation Society of America (ITSA), in 2004 more than one-half of state DOTs use a loop detector spacing varying from one-half mile or one-third mile or less for urban areas.() Figure shows the configuration of loop detectors on freeways and expressways. Although closely spaced loop detectors are helpful to system operation, operation of such a system imposes a significant economic burden and enormous pressure on system maintenance, and loop detectors do not directly provide travel times and have limitations in capturing area wide traffic dynamics. With the proliferation of cell phone usage, loop detector data is no longer needed for incident detection. Thus, the effectiveness of using loop detector data to estimate reliable travel time has not been proved. In recent discussions on traffic counting and detector spacing, researchers wondered if the spacing of detectors could be extended, for example, to one mile or more. (2) The focal point of the debate is to determine the appropriate detector spacing needed for different applications while not to degenerate the quality of data. Researches within the field of travel time estimation are incorporating traffic data from other sources than loop detectors to get more accurate travel time estimates. Probe vehicles have shown their potential as a valuable real-time traffic data source. This kind of vehicle-based system is expected to provide high-reliable traffic information when there are enough sample sizes. 2

14 Figure. Typical Configuration of Loop Detectors on Freeways. Statement of the Problem The Ohio Department of Transportation s (ODOT) advanced regional traffic interactive management and information systems (ARTIMIS) center in the Cincinnati area is capable of computing estimated travel times for key freeway corridors using the accessibility of Inductive Loop Detector Data (ILDD). The estimated travel times are currently relayed to motorists using ARTIMIS website and dynamic message signs (DMS) located at key locations along the critical freeways. Given the extensive use of travel time information and the popularity of ILD data, there is a need to evaluate and investigate methods to estimate travel time from this data. Drivers are interested in knowing how good the posted travel time data is so that they can make right decisions of their next step trips. However, inductive loop detectors have their own problems, such as not responding at a certain time, under or over counting vehicles, and placing economic pressure on system maintenance. There is an accuracy issue when the traffic volume reaches 3

15 congestion. Therefore, maximizing the efficiency and capacity of existing loop detector systems is vital because of the continued increasing in traffic volume and the limited construction of new freeway facilities in urban, intercity and rural areas. There are different methods available to calculate travel time from loop detector data; the most popular among them are using the point speed values, such as average speed approach, midpoint approach, minimum speed approach and weighed point speed approach. However, it is known that the accuracy of these speed-based methods will reduce as the vehicle flow becomes higher. Therefore, there is a need for an economically viable and accurate method of estimating travel time form loop detector data based on the theory of traffic flow that can be used under dynamic traffic flow condition, especially under congested traffic flows..2 Research Objectives and Methodology This research explored the accuracy of the current midpoint and other travel time estimation methods for different traffic cases. The comparison is based on both simulation platform and ground truth (probe vehicles) level. For ground truth data collection, Plate Matching Surveys was conducted License to study the reliability of probe vehicle method. Sample size issue associated with probe vehicles was addressed in this thesis. Another aim of this research is to explore optimal loop detector spacing for travel time estimation. This includes finding a reliable method to estimate travel time from the detector data and determine how detector spacing would affect its ability to produce accurate results. This study is necessary and timely because of the widespread use of the loop detection system today. Starting with the typical /3 mile detector spacing, 4

16 spacing was increased by using fewer detectors to obtain data, and at the same time different data processing methods reported by other researchers were applied to obtain travel times. An analysis on a sixteen-mile stretch of I-75 in Cincinnati, Ohio is performed and both simulation and field testing were used to evaluate the travel time estimates. Preliminary results showed that the accuracy of the midpoint method and other travel time estimation methods will generate larger error for links with incidents and large detector spacing. Different detector spacings result in different levels of accuracy. Moreover, as the traffic congestion sets in, the estimation errors from the existing methods increased significantly. As an improvement, a formulation with an error correction term was later introduced to account for the effect of congestion. This term is based on continuum modeling which can handle queue propagation and dissipation during travel time estimation. In comparison, the improved method is able to make more accurate travel time estimation, and can handle a variety of traffic conditions including work zone and incident condition at larger spacings. The estimated travel times from the improved method matches closely with those measured by the floating cars; the differences are within 0%. Results from this study showed that larger detector spacing than commonly used /3 mile do not worsen the estimation results. Overall, it seems to show that the one-mile spacing scheme has outperformed the other tested alternatives. 5

17 .3 Organization of the Thesis In Chapter II, a literature review was conducted to explore travel time data collection and freeway travel time estimation methods. A literature review on the detector spacing effect on freeway travel time estimation was also conducted. In chapter III, different freeway travel time estimation methods were discussed in detail. Subsequently, a VISSIM simulation model was introduced to study midpoint method and improved midpoint method. A statistical analysis was presented on travel time estimation and the effect of detector spacing on the results was studied. Field data collection and analysis was conducted in Chapter IV. In the first section, GPS equipped probe vehicle method was introduced and speed comparison between GPS data and loop detector data was presented. Evaluation of freeway travel time estimation with field data using improved midpoint method was completed. In the second section, license plate matching surveys were conducted on freeway sections to study reliability of probe vehicles. Summary and analysis of the matched vehicle travel time was conducted and the sample size issue associated with probe vehicles was addressed. Finally, Chapter V presents the conclusion of the thesis and recommendations for further research. 6

18 CHAPTER II LITERATURE REVIEW 2. Introduction This chapter reviews the literature related to the various aspects of inductive loop detector data (ILDD) collection and its application in travel time estimation. The chapter starts with a summary of loop detector data collection technologies and their accuracy. It is followed by a review of the methods used on freeway travel time estimation. The issue of the relationship between detector spacing and freeway travel time estimates is reviewed next. The final section of this chapter discusses the deficiency in the existing freeway travel time estimation studies and the innovation of this thesis. 2.2 Inductive Loop Detector Data (ILDD) Collection Technologies Current vehicle detection is based predominantly on inductive loop detectors (ILDs) installed in the roadway subsurface. When properly installed and maintained, they can provide real-time data and a historical database against which to compare and evaluate more advanced sensor systems. Inductive loop detectors were first employed by transportation engineering as detection equipment in the 960s and have become the most widely used field detection devices since then. [3] They are comprised of copper wire loops carrying an electrical current embedded in a shallow saw cut in the road surface. 7

19 An electromagnetic field is thus established that is interrupted when a vehicle passes over the loop. A change in the electrical current passing through the loop will be measured to determine corresponding traffic data information. When properly installed and maintained, inductive loop detectors can provide real-time data and a historical database. The data supplied by conventional inductive loop detectors include vehicle presence, vehicle count (volume), and occupancy (percent time a detector is occupied by one vehicle). Although single loop detectors cannot directly measure speed, it can be estimated by dual loop detector systems (two-loop speed trap). Loop data are typically relayed to a centralized Traffic Management Center (TMC) for analysis. This kind of data usually are available many times per second, however, the data will commonly be accumulated and amplified at the pull box and then reported to TMC at interval of 20 or 30 or 60 seconds. Some TMCs archive the reported detector data, while others aggregate them to 5-minute intervals and archive them after a filtering process. As reported in some literatures, inductive loop detectors have some advantages, such as fast data processing, broad and 24-hour coverage and all weather working condition. However, they also have some disadvantages. For example, there is a need for extensive system maintenance and calibration which will usually cost approximately $50,000- $60,000 with an additional $2,000 per year. [4] The quality of the data recorded by the inductive loop detectors is affected by any malfunctions such as over count, miscount, and inconsistency between two adjacent loops. Moreover, the data collection is continuous, which will make the error get accumulated. Therefore, quality control procedures become necessary in terms of employing loop data into different applications. 8

20 Many studies have been conducted in terms of diagnosing and screening detector errors. In some simple procedures, it is considered that equipment redundancy provides sufficient information to cover gaps in missing data or missing data is simply reported as is and decisions are made without this data. For several TMCs, field equipments were tracked through the maintenance of data base and problems such as the average percent of failed sensors were reported. Some traffic management centers evaluate the accuracy of new types of sensors before widespread deployment. [5] Besides answering the question of whether or not the data is good or bad, Z. Wall and D.J. Dailey [6] proposed a time series alignment algorithm to detect and recover errors that were not detected when data was originally collected. The basic idea was to find a reference station that is correctly calibrated and then the stations adjacent to the reference station will be adjusted and then used as the reference station for the next adjacent station correction. Smith and Conklin [7] used time-of-day historical average lane distribution patterns with current available detector data to estimate missing detector data. W.A.M Weijermars [8] et al checked the basic quality of loop detectors based on minimum and maximum flow thresholds, whereas quality checks based on the conservation of vehicles implied that flow measurements have to be consistent between upstream and monitoring detectors within one intersection. 2.3 Freeway Travel Time Estimation (TTE) From ILDD Engineers and planners have used travel time and delay studies since the late 920s to evaluate transportation facilities and plan improvements. Nowadays, travel time information is becoming more important for applications ranging from congestion 9

21 measurement to real-time travel information. Accurate travel time information is critical to dynamic travel assignment, motorists guidance and detour, effectiveness of transportation systems measurements, and determination of the location of traffic congestion area.[9] Therefore, providing drivers and TMCs with real-time, good-quality traveling information is becoming increasingly important, especially in many advanced traveler information system (ATIS) functions. There are various travel time estimation (TTE) methods based on different traffic data collection methods. Today, real time data can be gathered from microwave radar, automatic vehicle tag matching, video detection, license plate matching, and most commonly, inductive loop detectors (ILD). This thesis focused on freeway travel time estimation method from inductive loop detector data (ILDD) and evaluated it with probe vehicle data. As discussed before, single loops can directly measure volume and occupancy. Speed can be measured directly with dual loops or by calculation from single loop measurements with an estimate of vehicle length. Most jurisdictions apply the speed at a detector station (a location where loops cover all lanes) to a wider area---- typically half the distance to the next detector. While there is no guarantee that this represents the average speed over this segment, it is considered a reasonable estimate and is widely used in many DOTs. The final report of evaluation of Freeway Travel Time Estimates [0] mentioned midpoint and modified midpoint method to estimate link travel times by Oregon DOT (ODOT). In midpoint method, an influence area is assigned to an individual detector based on the locations of the midpoints between that detector and the next stations upstream or downstream. Probe travel times were compared to estimates of 0

22 travel time using midpoint and modified midpoint method. The results showed that the variation of midpoint and modified midpoint method is higher than probe vehicles (ground truth data). In terms of percent difference between probe vehicle travel time and midpoint method travel time, although it is fine under free flow traffic, this value tends to be very high (up to 65%) when traffic is congested, for example, and incident happened. Because a dual loop detector station is capable of recording vehicle speed and arrival times at a single point in space, Coifman [] proposed a method applying congested signal velocity u c, vehicle velocity v j, headway h j to calculate the estimated travel time. The example showed in this study got the travel time estimation which can be within 0% of the true value. But most TMCs or operating agencies have to upgrade their hardware and/or software in the filed to estimate travel time based on individual vehicle headway measurements. Jun-Seok Oh et al [2] used an algorithm based on the concept of section density that can be easily obtained by observing in-and-out traffic flows between two point detector stations with some careful corrections for detector errors. Travel times estimated from the proposed method are compared to those of other methods via both simulated and real traffic data. From the evaluation, it showed that the point measurements can be used for travel time estimation under non-congested traffic conditions but tend to underestimate during congestion period. Alternatively, the proposed method using the section density was able to estimate the travel time within 5% error. This study confirms that under congested traffic, to simply apply loop detector point measurement to estimate freeway travel time is not sufficient. Chu et al [3] discussed how to use Adaptive Kalman Filter (AKF) to estimate travel time. This algorithm fused both loop detector data---- volume count, speed and real travel time data

23 from a small sample of probe vehicles to apply AKF getting travel time. From the evaluation results, the proposed AKF can better estimate travel time throughout the whole study period, especially during the congestion period. But the source of ground truth was not stated clearly which will make the results less convincing. From the detailed review of literature, it is clear that various TTE methods based on loop detectors have been developed by researchers. The results seem that all these methods work fine only at low to moderate volume conditions, where the variability in the flow is lower. The performance decreases with increased flow conditions, especially for congested traffic flow conditions where travel speed is impeded by the formation and dissipation of vehicle queues and, in the mean time, the loop detectors can only provide data from isolated local points. Thus there is a need to develop a comprehensive TTE method that can be used for estimating travel time under various traffic flow conditions. 2.4 Detector Spacing Issue and Freeway Travel Time Estimation Currently, many state DOTs use detector spacing of one-third mile or two-thirds mile in congested urban areas and one mile in areas with minor or no congestion area []. Heavy reliance on closely spaced loop detectors implies significant economic burden. On the other hand, large detector spacing can result in loss of information, which leads to unreliable traffic information, especially in periods of congestion. Recent studies [4] found that with the proliferation of cell phone usage and the advent of computer-aided 5 traveler information systems, it is not necessary to require a high density of vehicle detectors along the highway for incident detection. At the same time when other advanced detectors such as video and microwave detection systems can be used to 2

24 provide traffic volume, vehicle speed, lane occupancy, as well as queue length information, traffic engineers wonder if the detector spacing can be increased while not sacrificing the quality of data for travel time estimation. The problem of finding effective and economic detector spacing has not been resolved by the existing studies. Among them, Kothuri et al. [5] summarized the results of a comparative analysis between two travel time algorithms the midpoint algorithm using the collective speed and a piecewise speed accumulation method requiring individual vehicle detection. In nine pipeline sections ranging from 0.5 miles to.5 miles, the authors found large estimation errors (5~20%) at some sections regardless of the data processing method. They further concluded that the accuracy of estimates seem to depend on the formation of queue with respect to the detectors. Wouters and Chan [6] provided a statistical algorithm to compute the average travel time for any freeway journey by using loop detector data for all Dutch freeway sections from the previous two years. This study showed that it is possible to make travel time prediction by using historical loop data when there is no construction or incident. Iris Fujito et al [7] studied the effect of sensor spacing on performance measure calculations. By comparing Travel Time Index (TTI), the ratio of free flow speed over the captured speed from the sensors, they concluded that as more sensors were deleted (-mile versus 3-mile spacing), the TTI measure did not become worse. Since TTI directly influence travel time, this study suggested that it is possible to use a reduced amount of data from limited loop detectors while maintaining the overall quality. To ensure the quality of traveler information, it is necessary to use an effective travel time estimation method which can perform consistently in different traffic conditions. 3

25 Motivated by the need for further studies on this problem, the research in this thesis explores improvements in travel time estimation and looks into the possibility of using larger detector spacing than the existing practice. Specifically, two issues were investigated: how the estimation error is affected by the spacing in different traffic conditions, and if enhancements can be made in the methodology so that the errors can be kept within an acceptable range when large spacings are used. 2.5 Concluding Remarks The significant investigation and widely use of ILDD, the advantage and disadvantage of them, the errors associated with them, the available methods for correcting these errors, and freeway travel time estimation were reviewed in this chapter. From the discussion in this chapter, it is clear that despite recent development efforts in many other advanced traffic surveillance systems, inductive loop detectors are still the most widely used traffic detectors. Due to feature of the invalid archived loop detector data, some procedures are needed to diagnose the ILDD and have data quality control before these data get further used to provide information such as travel time estimates. This chapter also reviewed the literature on the estimation of freeway travel time from ILDD. One observation can be made is that different methods perform well under offpeak hours. However, the problem remains in congested traffic, where travel speed is impeded by the formation and dissipation of vehicle queues and some important information will not be caught by isolated local points. Therefore, it is necessary to use an effective travel time estimation method which can perform consistently in various traffic conditions. Also, the issue of the relationship between detector spacing and travel 4

26 time estimation has been brought up in this chapter. Because of the economic burden, pressure on system maintenance, and uncritical to incident detection, there is a trend to use limited number of loop detectors. Since little study has been done within this area, this thesis tried to explore optimal loop detector spacing for travel time estimation. The details of the analysis will be discussed in the following chapters. 5

27 CHAPTER III FREEWAY TRAVEL TIME ESTIMATION AND COMPUTER MODELING 3. Introduction Based on the discussion in Chapter 2, different methods reported for the estimation of travel time form ILDD can be divided into two categories which are spot speed methods and traffic flow theory method. 3.. Spot speed methods Spot speed methods are based on the assumption that speed can be assumed to be constant for the small distance between two adjacent detector stations (approximately one third mile to a half miles). Since the distance between the two detectors is known, the travel time is simply calculated as the distance divided by the speed. Besides the midpoint method mentioned in the last chapter, there are three other spot speed methods used widely. A brief description of each method is given in the following sections Average Speed Method The speed of the vehicles traveling on a segment between two adjacent detectors is assumed to be the average of the speeds measured by the detector stations. Therefore, the travel time for whole corridor is given as: 6

28 Travel Time (TT) = L M + 2u 2 ( um D + u M M ) / 2 L + 2u M M (3.) D M- : length between loop detector station M- and M. u M- : speed at loop detector station M-. u M : speed at loop detector station M. Figure 3. Illustration of Spot Speed Method Minimum Speed Method In this method, the minimum of the speeds reported by the two adjacent detectors is treated to be the speed of the vehicles traveling on the segment. Therefore, the travel time for whole corridor is: Travel Time (TT) = L D L M M M + + (3.2) 2u 2 umin 2u M u min : minimum speed of V M- and V M Distance/Volume Weighted Average Method Travel time on each segment between upstream and downstream detectors is weighted according to either the spacing from the midpoint to each detector or the volume measured by each detector. Distance Weighted Equation: 7

29 8 Travel Time (TT) = M M M i i i i i i i i i u L L L L u L L u L u L = = M M M i i i i i i i i i u L L L u u u L u L u L 2 ) ( = (3.3) Volume Weighted Equation: Travel Time (TT) = M M M i i i i i i i i i u L N N N u L N u L u L = = M M M i i i i i i i i i i i u L N N u u u N L u N L u L 2 ) ( = (3.4) Where, L i is section distance, u i is section speed, N i is section volume Midpoint Method In the midpoint method, each loop detector station on the freeway network has an influence area which is assigned to an individual detector based on the locations of the midpoints between that detector and the next station upstream or downstream. In the case that no downstream station exists, the detector s influence area is extended to the end of the segment of interest. Figure 3. shows the concept of this method. Consider a corridor with M segments, indexed by j=,,m. At time slice i=,,t, the segment travel time can be calculated by:

30 t ( i j), = v L j ( i, j) (3.5) Lj: the segment length; v (i,j): the uniform speed within the section (obtained from loop detectors at isolated single points along the freeway). As discussed in the last chapter, the advantage for spot speed methods is that they are easy to apply. However, the performance will decrease with increased flow conditions. The discrepancy between estimated travel time calculated based on the spot speed method and the actual travel time during peak periods can be very big. Also, if the spacing between detectors is larger than one third to half a mile, the detector station will fail to capture the realistic traffic congestion information occurring between detectors, which will make travel time estimation less accurate Traffic Flow Theory Method In section 2.3, many travel time studies listed are based on traffic flow theory, which can capture the dynamic characteristics of traffic (Coifman [], Jun-Seok Oh [2], and Chu [3]). In these methods, the conservation of vehicles principle is applied and there is a need to develop a comprehensive model that can be used for estimating travel time under varying traffic flow conditions. A detailed discussion of this model will be given in the next sections. Consider two traffic counting stations on a one-way link as shown in Figure 3.2. In this figure, there are no traffic sources or sinks between stations, and station 2 is downstream from station. 9

31 Station x Station 2 Figure 3.2 Illustration of Continuity Equation Let Ni be the number of cars passing station i during time t, qi is the flow (volume) passing station i during time t, x is the distance between stations, and t is the duration of simultaneous counting at station and 2. Suppose that N > N2. This implies that there must be a buildup of cars between station and station 2 based on the assumption that there is no traffic sink between the stations. Let (N2-N) = N, thus: N/ t=q (3.6) N2/ t=q2 (3.7) N/ t= q (3.8) Then, the build up of cars between stations during the period t will be (- q) ( t). Let k equal the increase of the concentration of cars between station and 2. Then, ( N 2 N) k= x (3.9) and, Under the assumption of conservation of cars, - N= ( k) ( x) (3.0) (- q) ( t)= ( k) ( x) (3.) and 20

32 q k + x t = 0 (3.2) If the traffic flow is continuous and the finite increments are allowed to become infinitesimal, then q k + x t = 0 (3.3) If there are entrances or exits between freeway sections, then function (3.3) will be: q k + x t = g( x, t) (3.4) Where, q (x, t)--- the flow rate at location x and time t (veh/hr); k (x, t)--- density at location x and time t (veh/mile); g --- a generation term, which is equal to the entering (positive) or exiting (negative) flow rate, or zero for a pipeline freeway section; t --- time; x --- space. The simple continuum model is very widely used because it is not as complicated to implement as the higher-order model. Earlier work by Michalopoulous and Yi [8, 9] had shown some initial success in modeling queue spillback and congestion dissipation with continuum models. Subsequently continuum modeling was tried to be incorporated into the travel time estimation process. Simple continuum model was applied in two steps. First, conventional applications of the conservation equation involved finding the state variables q and k at the given sections of the roadway in a fixed time step. In this thesis, however, since the focus was 2

33 travel time, the model was discretized in such a way that travel time for the roadway section was to be determined from the changing characteristics of traffic flow. Secondly, because congestion propagation and queue dissipation directly affect travel time, different numerical methods were explored in searching of one that can handle shockwave effectively. Through experimentation, the upwind scheme by Stegger and Warming [20] was selected which supports flux vector splitting to implement Equation (3.4). The resultant equation after numerical treatment has this form: t n+ j = t n j x x k g x t x q (, ) + n j x x k g x t x q (, ) n j (3.5) Where j --- indices in space; n --- time step, dependent of x ; x --- discretized segment size; for numerical stability, the selected no less than the product of free-flow speed and the size of time step; x should be q, k --- change in flow rate and vehicle density; g --- generation term; x and x --- a forward difference and a backward difference, respectively; The travel time going through a roadway segment can be calculated after implementing Equation (3.5). For example, for a pipeline section (g = 0) the travel time at the time of estimation can be determined by the following formulations. Case. Congestion propagation 22

34 23 n j n j n j n j n j n j n j n j n j n j j u x x k k k k k u u u u t + = ) 2( (u j > u j+ ) (3.6) Case 2. Queue Dissipation n j n j n j n j n j n j n j n j n j n j n j n j n j j u x x k u k u k u k u u u u u t + = (u j u j+ ) (3.7) Where, x is discretization distance (included in each segment depending on detector spacing), j is discretization index, n is time step, u is traffic speed, and k is vehicle density which can be obtained by using the flow rate and speed information. Comparing with Equation (3.5), Equation (3.6) or (3.7) represents an improved form of the midpoint method the second term in the equation is the simple midpoint form and the first term represents an adjustment to the congestion s effect on travel time. Equations (3.6) and (3.7) can be used to treat queue propagation and dissipation within a congested section. For other uncongested sections, where the convection property of traffic flow prevails, the middle-point method is used to obtain travel time estimates. Thus, the travel time for the entire corridor will consist of two parts, as shown below: = = = + =, s j s j b b b t u L TT (3.8) Where u is the midpoint speed and L is the length of the section; b covers all the sections in normal traffic conditions; s covers all the congested sections and each section includes one or several segments denoted by j. Because of the iterative nature of Equations (3.6) and (3.7), they can be easily implemented on a spreadsheet file once the boundary

35 conditions from the congested section are known. Using Equation (3.8), a number of test runs were conducted on pipeline sections and ramp junctions to calibrate the model. This work involves selecting proper sizes of the space segment and time step so that the calculated travel time is numerically stable in similar types of geometric configurations (Table 3.). After model calibration, Equation (3.8) was further applied to obtain new travel time estimations. Similarly, the work zone and incident scenarios in both low and high volumes were tested, and the travel time estimates were updated. Table 3. Calibration Results on Different Types of Roadway Sections Road configuration Percentage error (%) (a) on and off ramps (b) weaving region (c) pipe line Computer Modeling and Implementation Traffic simulation software has become very popular as a traffic analysis tool used in transportation analyses. This section will give a brief introduction of different kinds of traffic simulation packages followed by a detailed description of a freeway corridor simulation platform. Based on this platform, the accuracy of travel time estimation was tested under different traffic scenarios, and detectors were relocated to formulate various spacing schemes History and Background of Traffic Simulation Traffic simulation is commonly used in transportation engineering today. In an effort to manage traffic as efficiently as possible, traffic simulation has gained recognition as a 24

36 very useful tool for the design of improvements to urban freeway systems. Traffic simulation enables the engineer to predict the outcomes of a proposed change to the traffic system before it is implemented, evaluate the merits of competing designs, quantify traffic operations and assess traffic conditions for various design alternatives of transportation facilities. Traffic simulation models are generally classified into macroscopic, mesoscopic and microscopic models, depending on their level of modeling detail. [2] Macroscopic models (TRANSIT-7F, HCS) describe the traffic process with aggregate quantities, such as flow and density. They cannot model the interactions of vehicles on design configurations. Microscopic models (CORSIM, SYNCHRO, VISSIM, PARAMICS, INTEGRATION) describe the behavior of the individual drivers as they react to their perceived environments. They essentially produce trajectories of vehicles as they move through the network. They have embedded algorithms and rules describing how vehicles move and interact, including acceleration, deceleration, lane changing and different driving behaviors. Mesoscopic models combine the properties of both microscopic and macroscopic simulation models. Mesoscopic models are somewhat less consistent than microscopic tools, but are superior to some other traffic analysis techniques. These models simulate individual vehicles, but describe their activities and interactions based on aggregate (macroscopic) relationships. A macroscopic model is often sufficient for the purpose of evaluating a proposed modification. These models tend to be easier to calibrate, since their parameters can be directly related to field data available from the existing sensor infrastructure. However, 25

37 since microscopic traffic simulation depends on a large number of inputs and parameters, the calibration process may become more difficult. Nowadays, microscopic traffic simulation tools have wider applications than macroscopic models because of their flexibility to represent spatial and temporal demand patterns, different driving behavior habits, OD matrix and detailed traffic control functions, and strategies. Among all the different kinds of microscopic traffic simulation models, VISSIM is more commonly and widely used than others. This thesis includes a case study with VISSIM simulation; therefore, an introduction of this software will be given in the following section Introduction of VISSIM (Urban Traffic Simulator) VISSIM is a microsimulation program developed at the University of Karlsruhe, Germany during the early 970s. VISSIM is a microscopic, behavior-based, multipurpose traffic simulation program and has become increasingly popular throughout the world. It offers a wide variety of urban and highway applications, integrating public and private transportation. Even complex traffic conditions are visualized in an unprecedented level of detail, providing realistic traffic models. The traffic flow model of VISSIM is a discrete, stochastic, time step based microscopic model, with driver-vehicle-units as single entities. The model is based on the continuous work of Wiedemann at the University of Karlsruhe, and further calibrated and validated by PTV AG. Car-following and lane changing together form the traffic flow model, being the kernel of VISSIM. 26

38 VISSIM can be applied as a useful tool in a variety of transportation problem settings. The following list provides some outstanding features of VISSIM [22]: Easy network editing VISSIM provides the users with a convenient and intuitive network editor. It creates and edits traffic networks based on background images supporting many different formats. In VISSIM, users can define highly exact positioning of numerous network elements and import the network topology and fixed time signal control from SYNCHRO. Vehicle behavior modeling VISSIM utilizes the Wiedemann 99 car-following model for freeway travel based on the work of R. Wiedemann. This model contains ten modifiable car-following parameters that classify the reactions of drivers in one of four driving modes: free driving, approaching, following or decelerating. A Lane-changing behavior model has also been built in. Urban and regional traffic operations and control It has a wide range of applications such as non-signalized intersections, conflicting movements and coordinated and actuated traffic signals. There are built in fixed-time signal control and NEMA controller, and users can develop any other type of vehicleactuated signal controller that can be coded with VAP, the C-like traffic control language. Numerous analysis options VISSIM provides a wide range of customizable evaluation such as number of vehicles, average speed, travel time, delay time and length of traffic queue spillback. The most recent patch also includes the Node analysis for users to evaluate the LOS and delays for different intersections. 27

39 A variety of animation capabilities In VISSIM, users can visualize the vehicle movements either in 2D or 3D which will offer clearer and more descriptive presentations. Based on the discussion above, VISSIM convincingly shows efficiency in performing traffic operation analysis. In this thesis, VISSIM has been chosen as the simulation platform of travel time estimation study Methodology and Data Analysis In this study, the VISSIM program was used to set up a simulation platform on which it is convenient to create different traffic scenarios and relocate loop detectors to formulate various spacing schemes. The test site selected for the study was the southbound roadway of I-75 in Ohio from the I-275 interchange to the Ohio River. This is a 6-mile stretch of freeway in which some parts sustain heavy congestion during the day. The network contains 2-lane, 3-lane, 4-lane, and 5-lane sections with roughly 7 ramps and the speed limit changes from 45 to 55 miles per hour. Figure 3.3 A shows the freeway network (highlighted in bold line from A to B). This freeway stretch is implemented in the simulation environment to create a test bed for evaluating travel time estimation alternatives. (Figure 3.3 B) 28

40 Figure 3.3 A Freeway Links of I-75, I-7 and I-275 in Cincinnati (from Figure 3.3 B VISSIM Simulation of Entire I-75 Network 29

41 As the first step, a detector spacing of one-third mile was utilized to evaluate the effectiveness of the midpoint method according to Equation (3.5). [Currently, loop detectors on I-75 in Cincinnati are installed at approximate one-third mile spacing.] After placing such detectors in the test bed, the whole corridor was divided into many short segments. Local detector speeds in 30-second intervals and calculated travel times for each segment between the mid points of adjacent detectors were obtained. The averaged piecewise travel times were then added together to get the estimated overall corridor travel time, which was compared with the overall corridor travel time obtained directly from the model. Three types of traffic conditions were considered in the initial test: normal flow condition; incident condition for short duration (blocking two lanes at a four-lane section); work zone condition (closing one lane at a three-lane section). In the incident case, where two lanes were blocked for a 30 minutes, leaving only the other two lanes to release vehicles, a long queue spillback extended upstream into a three-lane section. The work zone case modeled a long-term lane closure condition. One lane of a three-lane section was closed for 2600 feet, resulting in few problems in light traffic but minor slowdowns in heavy flows. For each traffic condition, three volume levels were considered: low (400 veh/hr/lane~600 veh/hr/lane), medium (600 veh/hr/lane~200 veh/hr/lane), and high (200 veh/hr/lane~800 veh/hr/lane). Therefore, a total of 9 runs were performed for the one-third mile detector spacing situation. For each case, five runs with different run seeds were conducted to reduce random errors. Results from the tests showed that when traffic volume is low, the difference in travel times between the midpoint algorithm and the model is small; in congested flows, the 30

42 errors increased dramatically. Figure 3.4 shows a sample of the errors under the normal traffic condition. 8 7 Percentage Error (%) low volume medium volume Traffic Volume high volume Figure 3.4 Percentage Errors at Different Traffic Volume Levels For the same traffic conditions, detector spacing was changed to investigate its effect on travel time estimation. By using fewer detectors an average detector spacing of 2/3 miles, mile, 4/3 miles, 5/3 miles, and 2 miles was developed. Although spacing greater than 2 miles have been tested, results from those tests are very poor and inconsistent, and thus they are not included in the thesis. In each discussed spacing alternative discussed, the detectors are selectively located to cover as many ramp junctions and interchanges as possible. For each combination of detector spacing and traffic scenario a total of five runs for every volume level were conducted and a total of 2076 data records at 30-sec 3

43 intervals were collected and analyzed. The absolute error (AE) and relative percentage error (PE) were used to summarize the results, as: AE= Pi - Oi (3.8) Pi - Oi PE= *00% Oi (3.9) Where, Pi--- estimated corridor travel time for a certain detector spacing i; Oi--- corridor travel time reported from the output file for the same detector spacing i; i--- detector spacing index, including /3, 2/3,, 4/3, 5/3 or 2. The permitted relative error is set according to the desired accuracy of the travel time estimation. The value of the travel time estimate will help determine the desired accuracy. Existing study [23] suggests that an error of 3 minutes is acceptable if the corridor travel time is near 30 minutes; for longer trips, the tolerance level is considered to be 0%~5%. Figures 3.5 and 3.6 show the mean value of absolute errors and percentage errors when different detector spacings under different traffic scenarios are tested. In Figure 3.5, the errors in low volume conditions are all within 2 minutes (under 0% of total trip time) regardless of the traffic condition and the detector spacing. In general, as the detector spacing increases, the errors also increase but the changes are gradual. The incident case under high volume produced the largest errors (6~8 minutes or 2%~25% of the total trip time). As Figure 3.5 shows only the mean value, further checking on the data revealed that some errors in the incident case reached over 35%. Figure 3.6 shows the error range at 95% confidence interval in each case, where it can be seen that high traffic volumes 32

44 led to much larger errors in general, and the incident condition caused the largest error variations due to the magnitude of the sudden impact on the traffic flow. The results for the above tests were further analyzed in the One-Sample t-test. The power of this test is greater than a simple test of the means. The One-Sample t-test considers the alternative hypothesis that the value of relative error is less than the value of acceptable error, which is chosen as 5% according to the previous discussion. All test results are presented in Table 3.2, where the One-Sample t-test is displayed as p-values. The upper boundaries of the confidence intervals were calculated using a significance level of 95% (a=0.05). The values in bold characters represent situations where the estimation error is equal to or greater than the acceptable error. The high p- values represent greater imprecision of the estimated travel time. It can be seen from Table 3.2 that high p-values occurred in all cases involving an incident and some cases under work zone condition at high traffic volumes. In summary, the results from Figures 3.4, 3.5, 3.6 and Table 3.2 suggest that the midpoint algorithm is not a reliable method for travel time estimation under heavy traffic volume or in abnormal traffic conditions. Clearly, by simply using data from detectors located in isolated roadway points it was not able to make good travel time estimates in congested flows regardless of the detector spacing. For congested traffic flow conditions where the variability in the flow is higher, the travel speed is impeded by the formation and dissipation of vehicle queues; therefore, it is insufficient to only have speed data from isolated local detector stations. Improvements must be made on the data processing method. A weighted averaging approach was tested which uses data from two adjacent detectors (upstream and downstream detectors) weighted according to either the spacing from the midpoint to 33

45 each detector or the volume measured by each detector. This method has been introduced in section % Absolute Error (min) % 7.88% 7.4% 8.06% 5.24% 3.42% % /3 mile.28%.53%.20%.8% 2/3 mile mile /3 mile 2/3 mile 2 mile detector spacing Variable work zone condition(corridor travle time 8.20min) incident condition (corridor travel time 9.86min) (a) Low Volume Absolute Error (min) % % /3 mile 26.86% 25.6% 23.47% 2.78% 2.78% 6.58% 2.59%.35% 9.73% 8.22% 2/3 mile mile /3 mile 2/3 mile 2 mile detector spacing Variable work zone condition (corridor travel time 24.28min) incident condition (corridor travel time 29.45min) (b) High Volume Figure 3.5 Absolute Mean Error from Midpoint Algorithm 34

46 Percentage Error in Normal Flow 95% CI for the Mean Percentage Error in Normal Flow 95% CI for the Mean Percentage Error (%) Percentage Error (%) /3 mile 2/3 mile mile 4/3 mile 5/3 mile 2 mile /3 mile 2/3 mile mile 4/3 mile 5/3 mile 2 mile (i) Low Volume (ii) High Volume (a) Normal Flow Condition Percentage Error in Incident Condition 95% CI for the Mean Percentage Error in Incident Condition 95% CI for the Mean Percentage Error (%) Percentage Error (%) /3 mile 2/3 mile mile 4/3 mile 5/3 mile 2 mile /3 mile 2/3 mile mile 4/3 mile 5/3 mile 2 mile (i) Low Volume (ii) High Volume (b) Incident Condition Percentage Error in Work Zone Condition 95% CI for the Mean Percentage Error in Work Zone Condition 95% CI for the Mean Percentage Error (%) Percentage Error (%) /3 mile 2/3 mile mile 4/3 mile 5/3 mile 2 mile /3 mile 2/3 mile mile 4/3 mile 5/3 mile 2 mile (i) Low Volume (ii) High Volume (c) Work Zone Condition Figure 3.6 Summary of Percentage Errors 35

47 Table 3.2 One-Sample t-test P-Values, 95% Upper Confidence Bound for Midpoint Method (α=0.05) High Volume Condition p-value error, minutes (percentage error) Normal flow condition p-value Low Volume error, minutes (percentage error) /3 mile (4%) (.%) 2/3 mile (6%) (.2%) mile (6%) (.3%) 4/ (.6%) (.6%) miles 5/3 miles (8.5%) (4.5%) 2 miles (0.7%) (3.%) /3 mile (39.5%) (9.6%) Incident Condition Work Zone Condition 2/3 mile (4.0%) (9.3%) mile (28.8%) (7.6%) 4/ (59.5%) (3.7%) miles 5/3 miles (53.9%) (8.5%) 2 miles (57.5%) (5.24%) /3 mile (8.2%) (0.7%) 2/3 mile (9.7%) (.3%) mile (8.2%) (.5%) 4/ (.3%) (.2%) miles 5/3 miles (2.6%) (.8%) 2 miles (6.6%) (3.4%) In summary, no obvious overall improvements were obtained in either the light or heavy flow conditions compared with simple mid-point method. For congested flows, the weights were manually set to the detectors according to real time condition acquired from the system, and changed their values depending on whether the congestion was before, after, or within the two detectors spacing. The testing showed that the weights must be reset frequently to address variations in the traffic conditions, but the changes do not 36

48 follow a traceable pattern that can be related to the level of traffic. In fact, in this way, improvements can only be obtained in a few separate cases but the results are not supported by consistency and the overall errors are still in the 20%~ 30% ranges in high volume conditions. Continued tests were conducted on the improved data processing methods based on continuum models. The results are shown in Figure 3.7 and Table 3.3. It can be found from those results that:. In general, as detector spacing increases the errors also increase; 2. The improved method leads to a reduction in the estimation errors under work zone conditions by 3%~5% (Figure 3.7 compared with Figure 3.5); 3. Similarly, the errors under the incident condition are reduced by 3 minutes, or 0% for the high volume cases; 4. When detector spacing increases to 4/3 miles and 2 miles, the errors are near or exceeding the 5% allowable level; 5. The spacing of mile appears to be associated with the lowest errors, within 5%~8%. The above results are further corroborated by the statistics in Table 3.3, where it can be seen that only the travel time estimates at the spacing of 5/3 miles are less trustworthy at the 95% significant level. All the other p-values are very low or equal to zero which indicates that the errors are acceptable at a=0.05 significance level. 37

49 4 Absolute mean error (min) % 0.78% /3 mile.45%.45% 2/3 mile 3.08% 2.68% 0.96%.27% 2.92% 2.38% 0.92%.23% mile /3 mile 2/3 mile 2 mile detector spacing Variable incident low volume (corridor travel time min) work zone low volume(corridor travel time 8.48min) (a) Low Volume Absolute mean error (min) % 5.68%.84% 0.87% 6.92% 4.79% 5.44% 8.53% 2.96% 7.46% 6.29% 2.0% /3 mile 2/3 mile mile /3 mile detector spacing 2/3 mile 2 mile Variable incident high volume work zone high volume (corridor travel time min) (corridor travel time 24.29min) (b) High volume Figure 3.7 Absolute Mean Error of Different Detector Spacing under Different Traffic Volume 38

50 Table 3.3 Comparison of One-Sample t-test P-Values, 95% Upper Confidence Bound of Percentage Error (%) for Improved Method (α=0.05) High Volume Low Volume Condition p- value error, minutes (percentage error) p- value error minutes (percentage error) 3 mile (3.0%) (0.9%) Incident Condition (26.0%) (.9%) 3 mile mile (27.0%) (.5%) (30.9%) (.2%) 3 miles (30.0%) (4.%) 3 miles 2 miles (33.%) (3.2%) Work Zone Condition 3 mile (9.8%) (.2%) (.8%) (.9%) 3 mile mile (9.6%) (.6%) (3.2%) (.7%) 3 miles (4.4%) (3.4%) 3 miles 2 miles (6.9%) (3.8%) 3.3 Concluding Remarks From the simulation results, the midpoint method cannot handle freeway travel time estimation under congested traffic conditions very well. Clearly, the improved midpoint method is very effective in reducing the estimation errors, and this helps increase the clarity and dependability of the results. The above results seem to suggest that a detector spacing under mile has insignificant effect on the estimated travel times (and the tests 39

51 show that such spacing can go even higher if one is willing to accept a larger error limit than 5%). In addition, since the study site includes 6 ramps in 7 miles, it is wondered whether it is coincidental that the one-mile spacing scheme outperforms others alternatives or whether it is so because the one-mile spacing represents approximately the average detector spacing within the corridor. With those findings in mind, a field test was further conducted to see if the results would be consistent when travel times directly measured from the field are used in the comparisons. 40

52 CHAPTER IV FREEWAY TRAVEL TIME STUDY WITH FIELD DATA 4. General Description of Field Data Collection While travel time data can be obtained through various sources such as loop detectors, microwave detectors, radar, etc., it is unrealistic that the whole network is completely covered by such data collection devices. With the development of wireless communication and satellite technology, probe vehicles as detectors are considered an important source of real-time travel information for a variety of intelligent transportation system applications. Travel time data can be collected continuously during study periods using this method. Probe vehicle travel time data are collected within traffic streams; therefore the data will not be influenced by the experimenter or driver. There are three types of probe vehicle data collection systems which are commonly used in transportation studies, such as Signpost-Based Automatic Vehicle Location (AVL), Automatic Vehicle Identification (AVL), and Ground-Based Radio Navigation. In recent years, new applications that use cellular geolocation and Global Positioning System (GPS) have been introduced into probe vehicle data collection systems. In this thesis, we are more interested in probe vehicles equipped with GPS unit technique. Probe vehicles equipped with GPS devices can provide the users with real-time position data in latitude-longitude pairs, travel times, distances, routes and vehicle speeds 4

53 at highly accurate levels. By using these accurate GPS spatial and temporal data, transportation engineers can easily calculate the basic traffic data, such as travel time, travel speed and travel distance, etc. GPS technology offers low cost, combined with high location accuracy. However, GPS data files tend to have huge numbers of records, especially if data is collected at short time intervals, for example every second. As a result, it is necessary to have formal procedures for linearly referencing, storing, and retrieving the GPS travel time and speed data. One way to aggregate the GPS data is to store the data in highway segments or links so that later segment travel time and speed data can be easily viewed by users. 4.2 Data Collection The improved travel time estimation method was further tested with field data from the 6-mile I-75 corridor. The tests involved comparing the loop detector data with the travel time obtained from the floating cars. Because of time and budget limitations, it was not able to create as many runs and test traffic scenarios for comparison as in the simulation based study. The field data collection was conducted by a joint team from The University of Akron and the University of Cincinnati. Eight people were assigned to the floating cars and each car was equipped with a HAiCOM GPS receiver and a HP ipaq data logger (Figure 4.) supplemented with the travel time for Windows software for easy data downloading. The receiver has a built-in magnetic mount which was mounted on the roof of the vehicle and connected to the handheld data logger. Continuous floating car data were collected from 3:30pm to 6:30pm on Monday, December 8, 2006 along the fixed study route of I-75 42

54 from I-275 to the Ohio River, both northbound and southbound. The probe vehicles were spaced at approximately equal time intervals in the traffic stream when they were dispatched. There is a driving technique when collecting traffic data using floating cars. In general, the drivers should follow the flow of traffic. If no traffic is present, the driver will follow the posted speed limit. In addition, loop detector and video data covering the entire roadway stretch from ARTIMIS Center for this time period were obtained. The loop detector data contain the average speed, volume and occupancy over all the lanes for every 5 minutes. Traffic volume was heavy in both the northbound and southbound directions. It was a typical weekday with good weather and no construction activities. Several locations along the corridor were experiencing slowdowns due to minor incidents and high traffic volume. GPS Receiver GPS Antenna Data Logger Figure 4. GPS Equipment 43

55 4.3 Data Processing The posted mile marker system along the test site was selected as the referencing network. Mile markers start at 0 on I-75 at the Ohio River and end at 7 at the interchange of I-75 and I-275, with individual markers every fifth mile along the entire route. Referring to these mile markers as reference points, the operators pressed the START RUN and END RUN buttons to get the data for the studied corridor. The GPS file was transferred from the HP ipaq data logger to a PC. PC Travel for Windows software developed by JAMAR Inc. was used and data was output in different study groups. Features (positions and attributes), corridor distance, travel time and average speed data were exported. (Figure 4.2) In Figure 4.2, the data has already been polled by adding Nodes according to the distance from reference mile markers. In doing this, a probe vehicle s data at every detector station can be obtained in order to compare GPS data with the loop detector data. Finally, PC Travel for Windows software can generate a data collection study summary, which includes all the information mentioned before, as well as speed/distance profiles of all runs. See Appendix for details. 44

56 Figure 4.2 Polled GPS Data The speed plotted from the GPS receiver was compared with the speed collected by the loop detector stations. The 5-minute speed data were used for the comparison with the GPS probe vehicle data. To do this comparison, the probe vehicle data at or very close to each detector station were averaged and then compared with that obtained form the corresponding loop detector. Figure 4.3 shows the results of the comparison. In Figure 4.3, it can be seen that the probe vehicle data are generally higher than loop detector data( (a) and (b)), but at some locations the probe vehicle data fluctuated and the speed was lower than the loop detector station at some time intervals ((c) and (d)). However, the fluctuation trend is the same between probe vehicle speed and the loop detector ones. This involves the issue of the reliability of probe vehicle technique and the sample size requirements associated with this method which will be discussed further in a later part of this thesis. 45

57 (a) Speed Profile at milemarker 5.6 (b) Speed Profile at milemaker 9.3 (c) Speed Profile at milemarker 2.4 (d) Speed Profile at milemarker 0.3 Figure 4.3 Speed comparison between GPS data and loop detector data 4.4 Evaluation of Freeway Travel Time Estimation with Field Data Using the field data, a test was conducted to compare the loop data based travel time with that of the floating cars. The existing /3 mile spacing for several runs was used and it is found that the travel time difference was in the range of 7%~25% and increasing the spacing caused this difference to be even greater. The test was continued by using the improved method with input from the loop detectors and video camera, and a detailed comparison with the floating car method was made, as shown in Figure 4.4. It can be seen that the difference is reduced to 7%~0% for detector spacing between /3 miles 46

58 and 4/3 miles. Again, the -mile spacing seems to outperform the others. When the spacing was increased to 3/2 miles and 2 miles, the errors jumped to 20%~30%. This is because at such large spacings it was not able to use the loop and video data to cover some ramp junctions with congestion. This indicates that detector spacing is very sensitive in travel time estimation for traffic in the real world, and further enhancement on the estimation method is needed if greater spacing is so desired. Nevertheless, the overall effectiveness of the improved method is clearly shown. 25 Percentage Difference (%) /3 mile 2/3 mile detector spacing mile /3 mile Figure 4.4 Travel time estimation errors with floating car data 4.5 Reliability of Probe Vehicle Method As discussed before, probe vehicle techniques is a real-time travel time data collection methods and become more popular, especially with the proliferation of the use of Global Positioning System (GPS). There have been some studies on the use of probe vehicles to 47

59 collect real-time traffic data [24-26]. The feasibility of using probe vehicles to collect real-time traffic information is very important and it is necessary to determine the number of vehicles that should be equipped as probes. Too few probe vehicles can provide erroneous or misleading data; too many probe vehicles will be expensive and labor consuming. Also, the field data results in 4.3 showed that there is discrepancy between probe vehicle data and loop detector data from which it is need to explore the reliability of the probe vehicle method and the same size requirement associated with it. Limited research has been conducted to determine the probe vehicle sample sizes and some findings and results are contradictive with each other. Mei and Steve [27, 28] studied the number of probe vehicles for freeway travel time estimation based on microscopic simulation. They found that the number of probe vehicles is strongly related to the distribution of network travel time and traffic volume. When traffic volume is light or high, the minimum required probe vehicle percentage will be higher than medium traffic volumes. Ashish et al [29] did research on frequency of probe reports and variance of travel time estimates. They claimed that travel time observations for each link are correlated and also found that under situations of high congestion the sample size of probes required would be smaller which is contradictive with the aforementioned two studies. Turner and Holdener [30] presented the Houston Experience in probe vehicle sample sizes for real-time information. In this paper, the authors used basic sample size equations to explore the sample size of probe vehicles. Based on the results, it was noted that congested freeways require a greater number of probe vehicles than uncongested freeways. The conclusion seems too general compared with other studies, and the validity 48

60 of the equation is questionable because it should depend on the distribution of travel time according to Mei and Steve Introduction of Sample Size Requirements and License Plate Matching Method In this thesis, the reliability of probe vehicle technique to collect travel time information was explored. In order to get large a sample size out of a traffic stream, the License Plate Matching Survey technique was employed. Details will be discussed in the following parts. License Plate matching technique provides a valuable source of data for traffic studies, especially for vehicle identification or tracking information. License plate matching has been applied to uses such as O-D matrix estimation, tracking the routes of individual vehicles, trip travel time calculations, weaving section analysis, and deriving traffic pattern information and other traffic studies. [3-34] Early license plate matching techniques involved placing roadside observers along a corridor or at the checkpoint of a study area to write down the digits of vehicle license plate numbers with a time stamp, followed by attempts to match the recorded digits to establish the path taken by the associated vehicle. However, collecting large samples of license plates in the field is difficult and transcription of license plates is very laborintensive (typically 0 hours per hour of data collection). More advanced license plate studies employed audiotape recorders, which allowed a spoken record of full license plates. In this case, high traffic volumes would overwhelm the observers. And glare, fatigue, or inattention would cause the loss of some license plate observations. [35] Introduction of video cameras or camcorders makes the process of license plate matching 49

61 much less labor intensive than before. If video cameras of adequate quality were properly set up, license plate numbers could be captured and transcribed into a computer in the office after the actual time of data collection. More advanced techniques involved video with character recognition which can automatically transcribe license plates and arrival times into a computer using computerized license plate character recognition. Generally speaking, although the accuracy of license plate reading is an issue for highspeed freeways, and skilled data collection personnel are required for conducting the experiment, the license plate matching technique is able to obtain a continuum of travel times during the data collection period. The travel times are from a large sample of vehicles, which is useful to understand the variability of travel times within the traffic stream. This thesis will employ pictures with manual transcription method to do license plate matching. It relies on digital cameras to collect license plates in the field and human personnel to transcribe the license plates into a computer in the office after the actual time of data collection Data Source of License Plate Matching Data Collection The source of this empirical license plate data for the study was collected by Transportation Lab in the University of Akron. The study site was a freeway corridor bounded by state route 8 at the Second Street Exit to the north and I-77 at mile marker 23.6 to the south. The whole corridor is 8 miles long and there are 7 ramps, 8 of which are off-ramps and 9 are on-ramps. The checkpoints are chosen at the boundary sites which are state route 8, Second Street Exit and I-77 at mile marker The License 50

62 Plate Data collection team included 4 people, two at each checkpoint. Each group had a digital camera and a tripod set up. The key point using a digital camera to capture license plate numbers is to have high shutter speed. According to the experience, shutter speed at /2500 or higher is enough. Continuous function was used when taking picture so that it was able to get clear license plate shots every 2 seconds. Pictures will be post-processed by manually transcripting data into an Excel Table, which contains as much information as possible from the pictures. (Figure 4.5) Morning peak periods (8:00 a.m. to 9:00 a.m.) were used for the study. For analysis both weekdays and weekends were considered. The data and other information are summarized in Table 4.. Figure 4.5 Spreadsheet form of License Plate Numbers 5

63 Table 4. Summary of Data Collection Parameters Freeway corridors: Interstate I-77 North at mile marker 23.6 to State Route 8 North at Second Street Exit. 8 off ramps, 9 on ramps, 8 miles. Time Periods: 2 weekdays in November 2007 Morning Peak Period: 8:00 a.m. to 9:00 a.m. weekend day in November 2007 Morning Peak Period: 8:20 a.m. to 8:50 a.m Data Analysis 0247 pictures were collected in total and 8399 license plate numbers were recognized. The overall rate of license plate recognition was 92.6%, and all unrecognized plate number took place only when vehicles were blocked by other ones traveling in adjacent lanes. An exact match was defined when each letter or number of a license plate obtained at a source checkpoint was also obtained at destination checkpoint. In order to eliminate other possibilities, a travel time threshold for the upper and lower limits has been set for the testbed. Vehicles with travel time of longer than 5 minutes and shorter than 5 minutes will be treated as a valid match. A computer macro program was developed to define exact matches and where at least five of the characters on the plates matched (Seven characters are most commonly found on Ohio plates). The match rate is around 4%, which equated to 584 matched vehicles in total. (Figure 4.6) 52

64 Figure 4.6 Matched Result of License Plate Summary and Analysis of the matched vehicle travel time data were based on 5-minute, 0 minute and 5-minute time periods. The analysis time interval of data is from 8:20 am. to 8:50 am.. Over the analysis time period, average travel times for all matched vehicles were computed, and, 2, 3, 4, 5,, 8 vehicles were randomly selected from the population as the fleet size of probe vehicles to get travel times and analyze sample size issue. Each fleet size, it will have 30 to 00 random repetitions according to different numbers of matched vehicles in that particular time interval. Figure 4.7, 4.8 and 4.9 describe the relationship between relative error of travel time and fleet size of probe vehicles. The relative error of travel time decreases as the number of probe vehicles increases. But when the fleet size of probe vehicles is 4 or 5, increasing the sample size will not cause a remarkable decrease in travel time relative error. This can be seen from the flat trend of the figures after 4 probe vehicles. Figure 4.7, 4.8 and 4.9 also show that when employing the same number of probe vehicles, errors for different time of the day, day of the week, and analysis period are different. It is because a different data set has different travel time variability. 53

65 percentage error (%) Thursday Data Tuesday Data Saturday Data number of probe vehicle (a) 8:25 a.m. to 8:30 a.m percentage error (%) Thursday Data Tuesday Data Saturday Data number of probe vehicle (b) 8:30 a.m. to 8:35 a.m. Figure 4.7 Percentage Errors under Different Number of Probe Vehicles for 5- Minute Analysis Period 54

66 percentage error (%) Thursday Data Tuesday Data Saturday Data number of probe vehicles (c) 8:35 a.m. to 8:40 a.m percentage error (%) Thursday Data Tuesday Data Saturday Data probe number (d) 8:40 a.m. to 8:45 a.m. Figure 4.7 Percentage Errors under Different Number of Probe Vehicles for 5- Minute Analysis Period (Continued) 55

67 percentage error (%) Thursday Data Tuesday Data Saturday Data number of probe vehicles (a) 8:25 a.m. to 8:35 a.m percentage error (%) Thursday Data Tuesday Data Saturday Data number of probe vehicle (b) 8:35 a.m. to 8:45 a.m Figure 4.8 Percentage Errors under Different Number of Probe Vehicles for 0- Minute Analysis Period 56

68 percentage error Thursday Data Tuesday Data Saturday Data number of probe vehicle Figure 4.9 Percentage Errors under Different Number of Probe Vehicles for 5- Minute Analysis Period According to statistics, standard deviation is one of the measures of variation which shows the dispersion of the values around the average: Standard deviation (SD) = n i= ( x i n x) 2 (4.) Notice that the standard deviation of data must always be understood in the context of the mean of the data. Therefore, coefficient of variation (CV) is also examined. It is defined as the ratio of the standard deviation SD to the mean µ: SD Coefficient of Variation C v = (4.2) µ 57

69 Coefficient of variation is often reported as a percentage (%) by multiplying the above calculation by 00 and is referred to as the relative standard deviation (RSD or RSD%). The standard deviation and coefficient of variation of travel time for different data sets was computed, and Table 4.2 shows the results. From table 4.2 and Figures 4.7, 4.8 and 4.9, it can be seen:. For time interval 8:25 to 8:30, travel time on Saturday has the highest standard deviation or coefficient of variation, and Tuesday has the lowest, so the percentage error of Saturday is the highest and Tuesday s data has the lowest using the same number of probe vehicles. The same relationship between standard deviation (coefficient of variation) and percentage error was also found in other analysis intervals. 2. For time interval 8:30 to 8:35, travel time standard deviation are the same for both Tuesday and Thursday and their coefficient of variation do not vary much. Therefore it is hard to say which data set has higher percentage error. And in fact, in Figure 4.7 (b) and Figure 4.8 (a), it shows big fluctuations. 3. Because standard deviation and coefficient of variance for a longer analysis period is higher than that of a shorter analysis period, the percentage error of the longer analysis period is also higher than that of the shorter analysis interval. For example, the 5-minute analysis period is higher than the 5-minute analysis period, and the percentage error of the5-minute interval is also higher than that of the 5-minute intervals. Generally speaking, the lower the standard deviation and coefficient of variance, the lower the corresponding percentage error will be using the same number of probe vehicles. It is because the standard deviation and coefficient of variance of travel time can reflect the reliability of probe vehicle travel time. When freeways are under free flow 58

70 levels which mean travel time of each vehicle varies much, the travel time obtained from probe vehicles will be less reliable. The influence of fleet size and number of repetitions were further studied on obtaining travel time using floating car method. Two-way ANOVA (analysis of variance) was used to explore the equality of mean relative error under different fleet size and number of repetitions. Under 95% confidence interval, two sets of hypotheses were tested:. Null Hypotheses, Ho: There is no difference in the mean relative error for different fleet size. Alternative Hypothesis, Ha: There is a difference in the mean relative error for different fleet size. 2. Null Hypotheses, Ho: There is no difference in the mean relative error for different numbers of repetition. Alternative Hypothesis, Ha: There is a difference in the mean relative error for different number of repetition. Analysis was still conducted under 5-minute, 0-minute and 5 minute interval. P- value was examined to determine whether to accept Ho or reject Ho. If p-value is smaller than significant level 0.05, then Ho has to be rejected and Ha will be accepted which means different fleet size and different number of repetitions will have significant influence on travel time measurement, vice versa. Table 4.3 shows the p-value of two-way ANOVA test for the two sets of hypotheses. Most p-values are smaller than significant level 0.05 therefore the fleet size and different number of repetitions have significant effect on collecting travel time. The numbers shown in bold indicate high p-values for fleet size which means for that time interval 59

71 there is no differences in the mean relative error for different fleet size. Generally speaking, according to Table 4.3, it can be concluded that different fleet size and repetition numbers will play an important role in the accuracy of probe vehicles travel time. Based on the previous discussion, there is a probe vehicle sample size (fleet size) issue associated with floating car method. Sample sizes are the minimum number of vehicles that are required to act as probes to estimate travel time or speed for a fixed statistical accuracy or acceptable relative error. According to previous sections and most existing real-time travel information systems operation, an acceptable relative error of ±0% was set. A 95% confidence interval was used to explore minimum required sample size under 0% acceptable error range. Confidence Intervals give us an estimate of the errors in the data and define the precision of the statistical estimates. The upper limit of the intervals is checked to determine based on whether the upper boundary exceeds 0% acceptable error. Table 4.4 shows the results. For 5-minute, 0-minute and 5-minute periods, 95% confidence level and 0% relative error, the sample sizes range from probe vehicle to 5 probe vehicles. Examining Table 4.4 and Table 4.2 together, it is found that minimum required sample size is related with travel time standard deviation and coefficient of variation. Basically speaking, large standard deviation will end up with large sample sizes. For example, for 5-minute interval analysis, because Saturday data has high standard deviation and coefficient of variance, the fleet size is larger than data of weekdays. For weekday data (Tuesday and Thursday), the sample sizes range from to 3 probe vehicles every 5 minute. Minimum required sample sizes are different under different analysis 60

72 time period as the standard deviation and coefficient of variance of travel time are different. The coefficient of variance values for 5-minute periods range from 4% to 20%, whereas the c.v. values for 0-minute periods and 5-minute periods are slight higher and range from 0% to 20%, 0% to 7% correspondingly. Consider about real application of floating car method, it is difficult to get ground truth data and it is why floating car method was needed to conduct. Therefore, instead of looking the relative errors of travel time, the travel times under each fleet size was studied. Previous two-way ANOVA analysis has already shown repetition number under each fleet size will influence the relative error of travel time. Therefore, repetition is considered as a factor which will affect the required minimum fleet size. For the same fleet size, 3 was chosen as the repetition numbers and conduct One-way ANOVA analysis (95% confidence level) on the factor of repetition on the travel time. The hypothesis here is: Null Hypotheses, Ho: There is no difference in the travel time between 3 groups of probe fleet. Alternative Hypothesis, Ha: There is a difference in the travel time between 3 groups of probe fleet. P-value is used to check the significant of the factor of repetition. If p-value is smaller than significant level 0.05, then Ho should be rejected and Ha was accepted which means the travel times between 3 groups of probe vehicles are different from each other at 95% significant level. In reality, the meaning of this is that under this fleet size, the travel time from the probe surveys may not reflect the real travel time because it varies between groups. Therefore the fleet size has to be increased until the variation between groups is 6

73 small enough to make the p-value greater than or equal By this mean, it can determine the minimum required sample size. The results are summarized in Table 4.5. Compared the numbers in Table 4.5 with Table 4.4, it is found that Table 4.5 is slightly stricter than Table 4.4. It is because the repetition factor method takes account of travel time variation more than confidence interval method. The required sample size is still range from to 5 probe vehicles. Sample sizes on weekend data is larger than that of weekdays for most analysis time intervals. The difference of sample size between 5-minute, 0-minute and 5-minute interval is not obvious compared with Table 4.4. Generally speaking, the fleet size of 3 or 4 probe vehicles will be good enough to conduct floating car method with reliable travel time. This result is close to what were found before when studying the relationship between relative error of travel time and fleet size of probe vehicles. In that analysis, it was found that the error lines tended to be flat from 4 probe vehicles. 5- minute Interval 0- minute Table 4.2 Standard Deviation and Coefficient of Variance of Travel Time SD of Travel Time (min) Coefficient of Variance (%) time interval Tuesday Thursday Saturday Tuesday Thursday Saturday 8:20 to 8: NA NA 8:25 to 8: :30 to 8: :35 to 8: :40 to 8: :25 to 8: Interval 8:35 to 8: minute Interval 8:25 to 8:

74 5- minute Interval 0- minute Table 4.3 P-value of two-way ANOVA table (95% confidence interval) p-value of two-way ANOVA Tuesday Thursday Saturday time interval fleet repetition fleet repetition fleet repetition size number size number size number 8:20 to 8: NA NA 8:25 to 8: :30 to 8: :35 to 8: :40 to 8: :25 to 8: Interval 8:35 to 8: minute Interval 8:25 to 8: Table 4.4 Sample Sizes for 5-minute, 0-minute and 5-minute Analysis Periods Minimum Required Sample Size (95% Confidence Interval, 0% relatvie error) 5- minute Interval 8:25 to 8:30 8:30 to 8:35 8:35 to 8:40 8:40 to 8:45 8:25 to 0-8:35 minute Interval 8:35 to8:45 5- minute Interval 8:25 to 8:40 Tuesday Thursday 2 Saturday 5 Tuesday Thursday 2 Saturday 2 Tuesday 2 Thursday 2 Saturday 4 Tuesday 2 Thursday 2 Saturday Tuesday Thursday Saturday 5 Tuesday 3 Thursday Saturday 2 Tuesday Thursday 5 Saturday 2 63

75 Table 4.5 Sample Sizes for 5-minute, 0-minute and 5-minute Analysis Periods Based on Repetition Factor Minimum Required Sample Size (Based on 3 Repetitions) 5- minute Interval 8:25 to 8:30 8:30 to 8:35 8:35 to 8:40 8:40 to 8:45 8:25 to 0-8:35 minute Interval 8:35 to8:45 5- minute Interval 8:25 to 8:40 Tuesday 2 Thursday 3 Saturday 5 Tuesday Thursday 2 Saturday 3 Tuesday 4 Thursday 3 Saturday 4 Tuesday 3 Thursday Saturday 2 Tuesday 2 Thursday 4 Saturday 3 Tuesday 5 Thursday 3 Saturday Tuesday 4 Thursday 3 Saturday Concluding Remarks In this chapter, a field data collection using floating car was conducted. The travel time directly measured with floating cars was compared with the estimated travel time from loop detectors. The purpose was to verify the improved midpoint method using field data so that it is convincible and applicable. The results consistently support the previous findings that using a moderately larger detector spacing than the existing /3 mile can be as effective in travel time estimation. This chapter also includes a study of the reliability of probe vehicles. In order to get large sample size, License Plate Matching Survey was conducted. Data was collected 64

76 over an 8 mile long freeway corridor. Summary and analysis of the matched vehicle travel time data were based on 5-minute, 0-minute and 5-minute periods. Preliminary results showed that the accuracy of floating car technique was related with the standard deviation of travel time and analysis period. Generally, the higher the standard deviation and coefficient of variance, the higher the error of the measured travel time will be. After studying 95% confidence interval of the data, the minimum sample size was explored. It is found that more probe vehicles are needed during weekend and different analysis time intervals will require different sample size. A further study on the sample size was conducted based on the difference in travel time between repetitions. The results under 3 repetitions showed that the fleet size of 3 or 4 probe vehicles will be adequate to get reliable travel time data for light to medium level traffic flow. 65

77 CHAPTER V CONCLUSION Most metropolitan areas rely on a large detector network to collect real-time traffic data in support of traveler information system in which travel time is an important information. The network usually involves the operation of hundreds or thousands of detectors at a uniformly fixed spacing and it is therefore very expensive to operate and maintain. Limited work has been done in the existing studies to investigate the impact of detector spacing on the overall travel time estimation and its economic significance. This thesis examined the effectiveness of freeway travel time estimation from the widely used midpoint method when different detector spacings are considered. A sixteen-mile section of I-75 in Cincinnati, Ohio is used as a testbed where alternative detectors are included each time to define a separate spacing scheme in the evaluation. A wide range of traffic conditions are considered, including work zone and incident and recurrent congestion. By comparing the estimated travel time from point detectors and with the corridor traffic time in a number of testing cases, it was found that the estimation errors are relatively stable if detector spacing is moderately increased. The simple midpoint method works at an error level of 20%~30%, and use of a congestion adjustment term in the method is able to reduce the estimation errors to a 0% level in incident conditions if the detector spacing is not to exceed mile. Preliminary results showed that the estimation errors 66

78 generally increase with detector spacing, and that the commonly used /3 mile spacing of speed detectors are not necessarily superior to larger spacings in case of congestion. Furthermore, the test was conducted in the field by comparing the estimated travel time with travel time directly measured with floating cars. The results consistently support the previous findings that using a moderately larger detector spacing than the existing /3- mile can be as effective in travel time estimation. In order to support and verify the field data collection method, License Plate Matching Survey was conducted to study the reliability of floating car technique Results showed that the accuracy of floating car technique is related with the standard deviation and coefficient of variation of travel time and analysis period. Required sample size problem has been studied also. Based on two different methods, it was found that sample size is related with travel time standard deviation and coefficient of variation. It was also found that the fleet size of 3 or 4 probe vehicles will be adequate to get reliable travel time data for light to medium level traffic flows. The analyses in this thesis have all indicated that using -mile spacing in I-75 testbed the estimated travel times match most closely to the corridor travel times. Since the average spacing of the ramps and interchanges in the testbed is also at one mile, it is wondered if these two facts are coincidental. However, further studies are needed to determine if there is a strong correlation between the ramp spacing and the best performing detector spacing in travel time estimation. Further data collection and testing are necessary to study the reliability of probe vehicle technique and the sample size issue associated with it. Data sets that cover from high-congested to free flow traffic conditions need to be obtained to study all the factors that will influence the minimum fleet size. 67

79 Although the scope of study is limited in this thesis, it is believed that this work will be beneficial for both engineers and researchers in travel time estimation and traveler information system area. The research has a great implication because, if proven, fewer detector stations will be needed and the long-term savings in system operation and maintenance costs will be immense. Given the high demand for time and resources for this type of studies, improved data acquisition methods (for example, by using locationbased technologies such as GPS, and cell phone station handovers) should help reduce cost and increase efficiency. 68

80 BIBLIOGRAPHY. ITS Survey Results: Spacing of Detectors 2004 Agency Summary Sarath Joshua. Freeway Management, Traffic Counting & Detector Spacing. ITS Technology Forum. 6 December Lawrence A. Klein. Sensor Technologies and Data Requirements for ITS. Boston: Artech House Benjamin Coifman. Distributed Surveillance and Control on Freeways. California PATH Research Report: UCB-ITS-PRR November, Shawn Turner. Defining and measuring traffic data quality--- white paper on recommended approaches. 83 rd Annual Meeting of the Transportation Research Board, Washington DC. January, Z.Wall, D.J.Daisley. An algorithm for the Detection and Correction of Errors in Archived Traffic Data. TRB 2003 Annual Meeting CD-ROM Smith, B L, Scherer, W T, Conklin, J H. Exploring Imputation Techniques for Missing Data in Transportation Management Systems. Transportation Research Record, Vol. 836/2003, pp W.A.M. Weijermars, E.C. Van Berkum. Detection of Invalid Loop Detector Data in Urban Areas. Transportation Research Record, Vol. 945, pp Xiaoyan Zhang, John Rice, Peter Bickel. Empirical Comparison of Travel Time Estimation Methods. California PATH Research Report: UCB-ITS-PRR Castle Rock Consultants Inc, Portland State University. Final Report of Evaluation of Freeway Travel Time Estimates

81 . Benjamin Coifman. Estimating Travel Times and Vehicle Trajectories on Freeways Using Dual Loop Detectors. Transportation Research-B, Jun-Seok Oh, R. Jayakrishnan, Will Recker. Section Travel Time Estimation from Point Detection Data.82 nd Annual Meeting of the Transportation Research Board, Washington, D.C., Lianyu Chu, Jun-Seok Oh, Will Recker. Adaptive Kalman Filter Based Freeway Travel Time Estimation. TRB Annual Meeting The Ohio Department of Transportation Central Office, HNTB Ohio, Inc. The Akron-Canton Freeway Management System Detailed Project Plan. January Sirisha M. Kothuri, Kristin Tufte, Soyoung Ahn and Robert L. Bertini. Development of an ITS Data Archive Application for Improving Freeway Travel Time Estimation. National Science Foundation under Grant March 20th, Jacorien A.A. Wouters, Kin-Fai Chan, Joost Kolkman and Rutger W. Kock. Customized Pre-trip Prediction of Freeway Travel Times for Road users. Transportation Research Board 84th Annual Meeting. January Iris Fujito, Rich Margiotta, Weimin Huang, William A. Perez. The Effect of Sensor Spacing on Performace Measure Calculations. TRB 2006 Annual Meeting CD-ROM. January Panos G. Michalopoulos, Ping Yi, Anastasios S. Lyrintzis. Development of an Improved High-order Continuum Traffic Flow Model. Transportation Research Record Anastasios S. Lyrintzis, Ping Yi, Panos G. Michalopoulos, Dimitrios E. Beskos.Advanced Continuum Traffic Flow Models For Congested Freeways. Journal of Transportation Engineering, Vol.20, No.3, May/ June, Stegger, J.L.,Warming, R.F. (98). Flux vector splitting of the inviscid gas dynamic equations with application to finite-difference methods. Computational Physics, 40(2), U.S. Department of Transportation, Federal Highway Administration. Types of Traffic Analysis Tools. April PTV America. VISSIM Manual

82 23. Brandy Hicks Meehan. Travel Times on Dynamic Message Signs. ITE Journal September Young, Stan. Real-Time Traffic Operations Data Using Vehicle Probe Technology, 2007 Mid-Continent Transportation Research Symposium Comert, Gurcan, Cetin, Mecit. Queue Length Estimation from Probe Vehicle Location: Undersaturated Conditions. Transportation Research Board 86th Annual Meeting Munehiro, Kazunori, Takahashi, Naoto, Asano, Motoki. Using Probe-Car Data to Analyze Winter Road Traffic Performance in the Urban Sapporo Area, 5th International Symposium on Highway Capacity and Quality of Service Mei Chen, Steven I.J.Chien. Determining the Number of Probe Vehicles for Freeway Travel Time Estimation Using Microscopic Simulation. Transportation Research Board 79th Annual Meeting, Washington, D.C , 28. Mei Chen, Steven I.J.Chien. Dynamic Freeway Travel Time Prediction Using Probe Vehicle Data: Link-based vs. Path-based. Transportation Research Board 80th Annual Meeting, Washington, D.C Ashish Sen, Piyushimita Thakuriah, Xia-Quon Zhu and Alan Karr. Frequency of probe reports and variance of travel time estimates. Journal of transportation engineering Shawn m. Turner, Douglas J. Holdener. Probe vehicle sample sizes for real-time information: the Houston Experience, Vehicle navigation & Information Systems Conference Proceedings, Seattle, Washington, July, 995, pp Asakura, Yasuo; Hato, Eiji; Kashiwadani, Masuo. Origin-destination matrices estimation model using automatic vehicle identification data and its application to the Han-Shin expressway network. Transportation, Vol. 27, Issue: 4, pp Texas Transportation Institute, Texas A&M University System. Travel Time Data Collection Handbook Chapter 4. March Shuldiner, P.W., S.A. D Agostino, and J.B. Woodson. Determining Detailed Origin-Destination and Travel Time Patterns Using Video and Machine Vision License Plate Matching. Transportation Research Record , pp Turner, S.M. Advanced Techniques for Travel Time Data Collection. In Transportation Research Record , pp

83 35. Rajeev Gupta, Jon D. Fricker, and David P. Moffett. Reduction of Video License Plate Data. Transportation Research Record 804.Paper No

84 APPENDIX SAMPLES OF GPS DATA 73

85 74

86 75

87 76

88 77

89 78

90 79

91 80