Road Traffic Estimation from Multiple GPS Data Using Incremental Weighted Update

Transcription

1 Road Traffic Estimation from Multiple GPS Data Using Incremental Weighted Update S. Sananmongkhonchai 1, P. Tangamchit 1, and P. Pongpaibool 2 1 King Mongkut s University of Technology Thonburi, Bangkok, Thailand, 11 2 National Electronics and Computer Technology Center, Pathumthani, s3522@st.kmutt.ac.th, poj.tan@kmutt.ac.th, panita@nectec.or.th Abstract In this study, we investigate the method to increase the accuracy of traffic estimation based on multiple GPS data from GPS-equipped vehicles in Bangkok area, Thailand. We propose an algorithm to predict an average speed of a road segment by combining hourly speed profiles of the road with real-time GPS velocity data. The algorithm utilizes incremental weighted update of average travel velocity from both sources. The velocity of data from GPS is used for update every time the data comes in, whereas the speed profiles are used for update when there is no data for a period of five minutes. The outputs are compared with a speed obtained with the position-based method. In this algorithm provides the low average speed error results around 8 km/h approximately. T I. INTRODUCTION HE traffic congestion in Bangkok increasingly leads to significant travel delays and inefficient use of gasoline. If drivers could know the congestion status before hand, they could avoid going through congested routes. Currently Bangkok drivers can obtain some traffic information from radio or T programs which report traffic condition by eyewitness volunteers and police. Such report is augmented by hundreds of closed-circuit cameras installed throughout the city. Unfortunately, the coverage of these systems is extremely limited due to high installation and maintenance costs. It is practically impossible to install traffic monitoring systems densely enough to cover the entire city s road network. Although eyewitness and police reports can cover more area, they are subjectively determined by human whose expertise, judgments, and experience can be vastly different and sometimes inconsistent. Current trend of traffic estimation focuses on the use of mobile probes. Mobile probes are sensors attached to cars that are driven through traffic. Mobile probes have advantages over fixed probes in term of high coverage area, flexibility, and low cost. Common forms of mobile probes are GPS receivers and cellular phones that are installed in cars. The position data from GPS can be used to estimate travel time from one location to another. To make the position data useful, there must be enough GPS-equipped vehicles in an area, and each vehicle must transmit the GPS data back to the data center at regular intervals. The data can then be processed and used to estimate the road traffic status. This study utilizes GPS data from a taxi company in Thailand. The company operates a taxi service in Bangkok metropolitan area. The company has around 4, GPSequipped taxis in operation. Position data from all taxis are periodically transmitted back to the data center via local amateur radio communication system. However, the bandwidth of the communication system is fairly limited. Therefore, there is no guarantee that the data will be transmitted successfully. Some data may be dropped if the communication channel is busy. Given such condition, the number of GPS data obtained on each road segment at a particular time interval is quite limit. We cannot simply compute an average speed of a road segment because it is not statistically meaningful. To solve this problem, we propose an algorithm to predict an average speed of a road segment by combining speed profiles of the road with real-time GPS data. The algorithm utilizes incremental weighted update of average travel velocity with the speed profiles and online velocity data obtained from the GPS. This paper is organized as follows. Section II discusses related work. Section III the overview of our system is described. We describe how the data and our algorithm work in Section I. The algorithm result and accuracy is showed in Section. Finally, Section I provides the conclusion. II. RELATED WORK In most traffic estimation studies involving probe vehicles, a single GPS-equipped probe vehicle is used to estimate traffic condition and travel time of a road segment [1], [2]. The problem of using a single probe data is that the estimation accuracy depends highly on a driver s behavior. Other works that employs data from multiple probe vehicles include [6], [9]. Previous techniques for traffic estimation involve learning algorithms such as fuzzy logic [1], [5] and neural networks [7]. In addition, Kalman filtering technique [3], [4] is used to predict the travel time. These are seemed to be complex method that is difficult to study. III. SYSTEM OERIEW In this study, we propose a simple yet effective method of identifying traffic condition. Our technique uses profiles of /8/$25. 8 IEEE

2 speed characteristics on the road. The speed profile is obtained by computing statistical average of speeds from the GPS hourly over the past. The average speed could identify traffic condition on the road in that time only. However, the speed profile alone is not suitable for real time traffic prediction, because it depends entirely on the past history. If there is newly added traffic data it could not have much effect on the speed average, even though it should have a high impact due to its recent occurrence. Therefore, we propose the incremental weight update algorithm to be used with traffic estimation. In our algorithm, the average speed profile is used in combination with online GPS data to estimate traffic status in form of average speed on a road. The estimated speed is a running variable that is updated regularly from the online data and the speed profile. The procedure of the algorithm is shown in Figure 1. In the data processing, there are two methods to estimate the speed characteristic on the road: velocity-based and position-based. The velocity-based method is a method that uses only instantaneous speed samples (i.e., spot speed) to calculate the average or mean speed. The position-based method takes position difference of the same probe vehicle at the sample time to calculate speed of that road section (i.e. space speed). The architectures of our algorithm are illustrated in figure 1. No No Timer Yes Speed Profile x W 2 GPS elocity Data x W 1 Incremental Weighted Update Estimated Speed GPS elocity Data Yes I. METHODOLOGY AND DATA A. Probe ehicles Data We obtained the GPS data from a taxi company in Bangkok, Thailand. The fleet has more than four thousand taxis equipped with GPS units. The GPS units communicate with the data center via short-wave radio. Each taxi is scheduled to transmit data back to the data center every 3.5 minutes. However, due to limited bandwidth in the shortwave radio communication, the GPS data is recorded at irregular intervals (6 minutes on average). The transmitted data consists of taxi ID, position (latitude, longitude), heading, speed, and timestamp. We obtain a data set collected during September 1-3, 6. B. Data Processing We use the GPS data of inbound Sathon Bridge on Wednesdays in September 6 to establish the traffic profiles in our algorithm. The reason that we select the Wednesday data is because it represents normal weekday traffic characteristics. The Sathon Bridge is chosen because of its heavy traffic jam condition. In addition, data from the bridge segment is reliable because taxis are unlikely to park or stop for passengers on the bridge. In our algorithm, we used two main methods in our algorithm. 1) Obtain Road Profiling: We considered the average of all speed samples from vehicles on the inbound Sathon Bridge section. We use speed samples from every Wednesday in September to calculate the hourly average speed profile, p(n), of the inbound Sathon Bridge section. We compute the average speed profile, p(n), as follows. p(n) th sum of all speed samples at n hour = number of all speed samples at n hour (1) We divide a day into 24 hour, the sample speed in each hour are calculated separately. In (1), p(n) is obtained from sum of all speed samples at n th hour divided by number of all speed samples at n th hour. The results of calculation are shown in Table I and figure 2. th Figure 1. Architectures of our algorithm Average Speed In figure 1, if a timestamp of the GPS data input is longer than a timer value over the timestamp of the previous GPS data, a speed profile that is multiplied by the weight W 2 would executed in the incremental weighted update part and provide a predicted speed. However, if a timestamp of the GPS data input is less than a timer value under the timestamp of the previous GPS data, a GPS speed data input that is multiplied by the weight W 1 would executed in the incremental weighted update part and provide a predicted speed n-th Hour Figure 2. The speed profile in each hour on Wednesdays.

3 TABLE I THE AERAGE SPEED, P(N) N th Hour Time (HH:MM:SS) Average Speed(km/h) 1 :: :59: :: 1:59: :: 2:59:59 3:: 3:59:59 4:: 4:59:59 5:: 5:59: :: 6:59: :: 7:59: :: 8:59: :: 9:59: :: 1:59: :: 11:59:59 12:: 12:59:59 13:: 13:59:59 14:: 14:59:59 15:: 15:59:59 16:: 16:59:59 17:: 17:59:59 18:: 18:59:59 19:: 19:59:59 :: :59:59 21:: 21:59:59 22:: 22:59:59 23:: 23:59: in figure 3. The speed characteristic is determined by two calculations: incremental weighted update algorithm and position-based methods. They provide spot speed and space speed characteristics of the road. The spot speed is compared with the space speed characteristic result to determine the average error. GPS Data Our Algorithm Method Spot Speed Compare Space Speed Figure 3. The diagram of our evaluation system. Average Error 2) Incremental weighted update algorithm: All speed samples on Wednesday 27 th are applied to update the speed profile. The speed profile is updated by each speed sample input and the previous speed profile value is used to update if there is no speed sample input during five minutes. The equation that we use to calculate a new speed profile value is showed as follows. nn = [w 1 s]+[(1- w 1) np] (2) nn = [w 2 p(n) ]+[(1- w 2) np] (3) In (2) and (3), nn is a new updated speed profile. The variable w 1 and w 2 are the weight values. In (2), s is a speed sample of GPS-equipped taxi. In (3), p(n) is a speed profile in each hour which is weighted by w 2. A new speed profile, nn, is calculated by providing weight in (2) and (3); the weight is distributed over a region from to 1. The variable s in (2) is the speed sample of GPS-equipped taxi which is weighted by w 1. np is a previous updated speed profile of the road. np in (2) and (3) are weighted by 1-w 1 and 1- w 2 respectively. In case there is no speed sample input from GPS data during five minutes, the p(n) that is the speed profile value in the time section that five minutes is added. p(n) is used to updated a new speed profile of the road in (3).. EXPERIMENT AND RESULTS In our system, the accuracy of the incremental weighted update algorithm is evaluated against that of the positionbased method. Therefore, the evaluation system is illustrated A. Method The position samples that we obtain from GPS data on Wednesday 27 th are used in a similar manner as in the velocity-based method. Only taxis that send the position information more than one sample per hour on the Sathon Bridge are selected to calculate the space speed. Due to GPS data that we obtained is irregular intervals, this is impossible to match the GPS position exactly at the beginning and the ending points of road section. Therefore, all sample data that is in the road section would be available. The position of taxi could be estimated in term of distance in kilometer unit, we used (4) to compute a net distance D net in kilometer. The number 111 is the approximated value of the distance in kilometer unit of one geographical degree. 2 2 D net = (lat1 - lat 2) + (lon1 - lon 2) 111 (4) The position information in GPS data consist of latitude and longitude values. The lat 1 and lat 2 are latitude value; lon 1 and lon 2 are longitude value. The variable lat 1 and lat 2 are origin and destination position in term of latitude of the probe vehicle respectively. The variable lon 1 and lon 2 are origin and destination position in term of longitude of the probe vehicle respectively. T is the time difference between two different positions by the same taxi identification number. We compute the space speed ss using (5). D net ss = (5) ΔT

4 A timestamp of the origin position samples is defined as a timestamp of space speed results. B. Algorithm Evaluation To evaluate accuracy of our algorithm, we use the position-based method which is a more reliable method to estimate the speed on a road segment. The space speed from the position-based method is compared with the spot speed from the incremental weighted algorithm. We compute the average error of both speed results by using (6). sum of the difference of all both speed Average Error = number of all space speed data In this experiment, we investigate how the algorithm works with different weights, w 1 and w 2 values. We varied the weights between to 1 for both w 1 and w 2. Figure 4 shows spot speed and space speed comparison when w 1 is and w 2 is 1. The average error in this case is 1.34 km/h. The graph of spot speed looks like the hourly speed profile in figure 2. This is because we rely on the profile data 1 percents without considering the GPS velocity data (6) If we adjust the weight w 1 and w 2 to high and low,.9 and.1 respectively, and in case of low and high, and 1 respectively as shown in figure 6 and 4. These combinations provide the average error of 8.96 and 1.34 km/h. Next we adjust the weight w 1 and w 2 to the same value of.5, this combination provides the average error of 8.66 km/h as shown in figure 7. After experimenting with different weight combinations, as shown in Table II, it turns out that the weight combination of w 1 =.5 and w 2 =.1 as shown in figure 5 yields the lowest error. TABLE II AERAGE SPEED ERROR W 1 W 2 Average Error(km/h) Figure 4. The spot speed compare with the space speed using w 1 = and w 2 = Next we adjust the weight w 1 to.5 and w 2 to.1. This weight combination provides the average error of 8.63 km/h as shown in figure Figure 6. The spot speed compare with the space speed using w 1 =.9 and w 2 = Figure 5. The spot speed compare with the space speed using w 1 =.5 and w 2 = Figure 7. The spot speed compare with the space speed using w 1 =.5 and w 2 =.5.

5 Based on the algorithm, the value of w 2 should be small compared to w1 because the speed profile is used for updating every 5 minutes. It will eventually make the estimated speed converge to the profile if there is no addition online data. For every new traffic data entered, the speed will be updated only once. Therefore, the weight value for w1 should be high to make the predicted speed jump toward the new velocity data because of its recent occurrence. I. CONCLUSIONS In this paper, we propose an algorithm to predict an average speed of a road segment by combining speed profiles of the road with real-time GPS data. The algorithm utilizes incremental weighted update of average travel velocity. Results show that the proposed algorithm yields more accurate speed estimation on a road segment than the statistical average speed method. The incremental weighted update algorithm provides low average speed error around 8 km/h approximately. REFERENCES [1] Y. Li and M. McDonald, Link travel time estimation using single GPS equipped probe vehicle, in IEEE Intelligent Transportation System, September 2, pp [2] W. Pattara-atikom, P. Pongpaibool and S. Thajchayapong, Estimating road traffic congestion using vehicle velocity, in IEEE ITS Telecommunications Proceeding, September 6, pp [3] L. Chu, J. S. Oh and W. Recker, Adaptive Kalman filter based travel time estimation, in TRB Annual Meeting, 5. [4] J. S. Yang, Travel time prediction using the GPS test vehicle and Kalman filtering technique, in American Control Conference, June 5, pp [5] S. H. Lee, B. W. Lee and Y. K. Yang, Estimating of link speed using pattern classification of GPS probe car data, in ICCSA, 6, pp [6] J. Yoon, B. Noble and M. Liu, Surface street traffic estimation, in MobiSys, 7, pp [7] W. Pattara-atikom and R. Peachavanich, Estimating road traffic congestion from cell dwell time using neural network, in IEEE Conference on Intelligent Transportation Systems (ITSC 7), Seattle, USA, Sep 7. [8] S. Thajchayapong, W. Pattara-atikom, N. Chadil and C. Mitrpant, Enhanced detection of road traffic congestion, in IEEE Intelligent - Transportation Systems Conference, September 17-, 6, pp [9] S. Xiaohui, X. Jianping, Z. Jung, Z. Lei and L. Weiye, Application of dynamic traffic flow map by using real time GPS data equipped vehicles, in IEEE Conference on ITS Telecommunications Proceedings, 6, pp