Article begins on next page

Transcription

1 Feasibility of Software-Based Duty Cycling of GPS for Trajectory-Based Services Rutgers University has made this article freely available. Please share how this access benefits you. Your story matters. [ This work is an ACCEPTED MANUSCRIPT (AM) This is the author's manuscript for a work that has been accepted for publication. Changes resulting from the publishing process, such as copyediting, final layout, and pagination, may not be reflected in this document. The publisher takes permanent responsibility for the work. Content and layout follow publisher's submission requirements. Citation for this version and the definitive version are shown below. Citation to Publisher Version: Li, Xiaohan, Yuan, Fengpeng & Lindqvist, Janne. (2016). Feasibility of Software-Based Duty Cycling of GPS for Trajectory-Based Services. reload=true&punumber= Citation to this Version: Li, Xiaohan, Yuan, Fengpeng & Lindqvist, Janne. (2016). Feasibility of Software-Based Duty Cycling of GPS for Trajectory-Based Services. Retrieved from doi: /t3vm4f56. Terms of Use: Copyright for scholarly resources published in RUcore is retained by the copyright holder. By virtue of its appearance in this open access medium, you are free to use this resource, with proper attribution, in educational and other non-commercial settings. Other uses, such as reproduction or republication, may require the permission of the copyright holder. Article begins on next page SOAR is a service of RUcore, the Rutgers University Community Repository RUcore is developed and maintained by Rutgers University Libraries

2 Feasibility of Software-Based Duty Cycling of GPS for Trajectory-Based Services Xiaohan Li, Fengpeng Yuan and Janne Lindqvist Rutgers University Abstract Energy-efficient localization is increasingly important for many types of smartphone apps. The research community has argued that fixed duty cycling of GPS is not a good choice for trajectory-based services concerning route accuracy. In this paper, we describe the design and implementation of a highly accurate map matching and path construction algorithm. Furthermore, with thorough field experiments, we show that fixed duty cycling of a smartphone GPS receiver is a feasible approach for trajectory-based services, and it can achieve considerable energy conservation without sacrificing much route accuracy. When increasing the GPS sampling period beyond 120 seconds, it saves at least 78% energy in comparison to continuous GPS sampling, while the loss of route accuracy tends to be stable around 23%. I. INTRODUCTION Location-based services are increasing with mobile apps. Since location information serves as an important context to many mobile applications, almost all of the smartphones today have integrated GPS chips. Some of the location based applications are satisfied with a single position fix. However, some of them require recording location traces over time. For example, multiple location fixes are needed to reconstruct a route a car has taken. We call this type of services as trajectorybased services. GPS sensing is well known for its high energy consumption. The rate by which the GPS is sampled (also called duty cycle) strongly affects energy consumption. Continuous sampling of the GPS can drain the battery of a smartphone within few hours. Thus, energy efficiency has an ever increasing importance in trajectory-based services for modern mobile devices. In order to realize energy efficient localization, the most obvious way is using duty cycling GPS receiver instead of performing continuous GPS sampling. However, a lower sampling rate may lead to the object s movement between sampling points as uncertain [1]. Several works [2], [3], [4], [5], [6], [7], [8] have also cast doubt on fixed duty cycling GPS concerning route accuracy or energy efficiency. However, with the knowledge of road information, we are still able to construct the route from a GPS trace even when the GPS sampling rate is low and GPS samples are inadequate. Thus, fixed duty cycling GPS is a feasible approach for efficient trajectory-based services. The challenge is to find out the tradeoff between the accuracy of the routes constructed from GPS traces with different sample rates and energy consumption. In this paper, we show that we can formulate the most energy-efficient period of duty cycling GPS, and guarantee that the accuracy is high enough to construct routes from the traces. This paper presents the following major contributions: We present the results of experiments of a real-life traffic dataset. These results show the tradeoffs between route construction accuracy and energy consumption for software-based fixed duty-cycling of GPS. In the general case, when the sample period exceeds 120 seconds, the mismatched fraction of route tends to be stable at 23% to 25%, while saving at least 78% of energy. As minor contributions we have designed and implemented an algorithm for map matching particularly suited for trajectory-based services. We also present a path exploration method that can construct the path from inadequate GPS samples. We note that although all our computation was done with OpenStreetMap (OSM) data, we illustrate the concepts in this paper with Google Maps data due to aesthetic reasons. II. OUR APPROACH Our work has two main goals: 1) to formulate the relationship between the accuracy of the path constructed from GPS traces and the sampling rates, and 2) to analyze the trade off between path construction accuracy with GPS energy consumption under different sampling rates. To realize these goals, we first collected over a hundred GPS traces. Second, we implemented an algorithm to map match the GPS traces to OpenStreetMap (OSM) [9] data and constructed paths from it. With the map matching algorithm, we can quantify the path construction accuracy with different sampling rates. Additionally, we analyzed the energy consumption regarding the route construction accuracy and sampling rate. A. Data Collection We have collected 123 traces with different route complexity and road network density in Central Jersey. In the process, each participant carried an Android phone with an application (Torque) that can log GPS locations and store them into a local database during driving. The traces collected are all under normal driving circumstances. The routes we collected data on are distributed in different areas of Central Jersey and contain both highways and local roads. The total distance of the traces is miles. The longest trace is miles, while the shortest trace is 1.78 miles. The sampling rate for the GPS update was once per second. B. Parsing the GPS Traces Trajectory-based services are usually associated with road information or maps. To construct paths from GPS traces, we use the road information provided by OSM. OSM is a

3 Fig. 1: Difficulty of mapping. The red nodes represent a route, and the black node is one sample from a GPS trace collected on this route. If we simply match the black node to its nearest OSM node, this will result in a mismatch. As we can see from the figure, the blue node is its nearest OSM node, but it does not reside in the route. So we need a better map matching algorithm. Map data c 2013 Google. collaborative project to create a free editable map of the whole world. Due to the emphasis of local knowledge and ground truth in the process of data collection, the project has a geographically diverse user base and the density of map data of OSM varies from area to area. OSM provides the following information: Nodes, Ways, Adjacencies, Lanes, Speed types and Bounds. Given the form of the OSM data associated with each road, a path is a sequence of nodes. Fig. 1 shows all the nodes along a route in OSM. As Fig. 1 indicates, the distances between two nodes on the map are different. A straight road is represented by a relatively lower density of nodes, while for a curvy road, there are relatively more nodes representing it. The uneven distribution of nodes in OSM, as well as the GPS error, imposes difficulty on map matching. For example, in Fig. 1, the route is represented by those red nodes. The black node is one sample from our GPS trace. Our task is to match the black node onto the route represented by the red nodes. The obvious algorithm for map matching is to match the GPS sample to its nearest node (blue node). However, as we can see, the nearest node (blue node) obviously does not reside on the correct route. Therefore, a map matching algorithm is needed. C. Map Matching Map matching is the procedure of aligning a sequence of observed user positions with the road network on a digital map. There are a number of studies on map matching [10], [11], [12], [13], [14], [15]. Based on the road information provided by OSM, we implemented a map matching algorithm to match the GPS samples to the data points in OSM. Our approach utilizes a backtracking and nodes connectivity check. When matching the GPS samples to the map, two conditions we should bear in mind: 1) the candidates we select must be within a certain range of the matching node. 2) the candidates must be reachable from the nodes that have already been matched. Appropriately setting up the range of the matching node will significantly reduce the searching time because many impossible candidates are rejected, and we do not have to do some unnecessary matching work. During map matching, it is inevitable that some nodes will be mismatched due to the uneven distribution of OSM data, GPS error, etc. So the main problem becomes how we can correct the mismatching in order to achieve high map matching accuracy. OSM has an Adjacency database, which provides the successor and predecessor of the current node. We can take advantage of this to check the connectivity (reachability) of a previous matched node and the current matched node. If we have exhausted all of the candidates of the current sample and do not find a good match, we know that a mismatching appears before the current node. Then we need to backtrack to the previous nodes to find the mismatching and rematch that node(s). The pseudocode for this algorithm is given in Algorithm 1. There are several details we do not show in the pseudocode. For instance, even when we backtrack to the first node of the trace and do not find a good match, we will enlarge the range of searching area, but this is still within some reasonable range according to the GPS accuracy. As shown in Fig. 2(a), the red nodes are the nodes we have already matched. The black node is a sample from the GPS trace that we want to match onto the map. First, we get a bunch of candidates that are close to the black node and color them blue. Then we check the connectivity from node 1 to each of the blue nodes. Both node 3 and node 4 are reachable from node 1 while node 4 is closer, we match the black node to node 4 as shown in Fig. 2(b) and proceed to the next node (6). Similarly, we get candidates of node 6 and check the connectivity from node 4 to them. Now, an error arises since none of the candidates can be reached by node 4. Thus, we know it s a wrong decision to match node 2 to node 4. Then we backtrack and rematch node 2, meanwhile we mark the unqualified candidates grey. We match node 2 to node 3 since it s the only candidate left, as indicated in Fig. 2(c). Then we continue to match node 6, as shown in Fig. 2(d). This time we can tell that node 7 is a good match of node 6 because it satisfies all the conditions. Because of the backtracking and connectivity check mechanism, if a matching error happens, we can always detect it and backtrack to correct it. Therefore, this algorithm can achieve high map matching accuracy. D. Path Construction After map matching, we obtain the OSM nodes corresponding to the GPS nodes, accordingly, we could get the OSM ways in which the OSM nodes reside. In OSM, a way is an ordered list of OSM nodes, having way ID as the primary key in OSM. With these matched OSM nodes, we can construct the user s path. There are two scenarios of constructing the paths. 1) Two nodes have the same way ID. 2) Two nodes have different way IDs. For the first scenario, we can simply connect these two nodes. However, we cannot connect them by a straight line, because the road is rarely straight. Instead, we add some OSM

4 (a) Get valid candidates (b) Error matching (c) Backtracking and rematching (d) Proceed to next node Fig. 2: Map Matching Process. (a). First we get the matching candidates of the black node, sort them according the distance, and color them blue. (b). We check the connectivity from the previous matched node(node 1) to the blue nodes, and choose the one which is nearest to the black node and reachable from the previous matched node, color it red (node 4). Similarly, we get the candidates of node 6, and get the candidates. (c). When we check the connectivity from node 4 to the candidates of node 6, an error arises. None of them can be reached by node 4, which is telling us that the previous match is a mismatch. Then we backtrack and rematch node 2 to node 3. Meanwhile, we mark the unqualified nodes grey. (d). After backtracking and rematching, the algorithm continues matching next node till the end of the trace. Map data c 2013 Google. Algorithm 1: Map Matching Input: A GPS trace, T, # of nodes in T, N, a Stack, stk, # of matched nodes, k Output: Matched Trace in OSM 1 s 0 = first sample of T ; 2 candidates = get candidates(s 0 ); 3 sort candidates; 4 foreach candidate in candidates do 5 stk.push(); 6 k=0; 7 foreach sample in T do 8 cur candids= get candidates(sample); 9 sort cur candidates; 10 foreach candidate Not grey in cur candids do 11 if reachable by prev matched node then 12 stk.push(); 13 k + +; 14 if k==n then Done; 15 ; 16 else Break; 17 ; 18 else if Not reachable by prev matched node && hasn extcandidate() then 19 Continue; 20 else // backtracking 21 stk.pop(); 22 k ; 23 Gray popped candidate; 24 Rematch previous sample; 25 end 26 end 27 end 28 end 29 return reverse order of stk; nodes between them along the road which can be fetched from the Adjacency database, as shown in Fig. 3(a) and 3(b). The number of points we add between these two nodes depends on the density of nodes along the road in the database. Then we can construct the path by connecting these nodes one by one. To construct the path under the second scenario, we cannot connect them using the above scenario, since the road network is much more complex. With the help of the Adjacencies database, we apply a modified Breadth First Search (BFS) algorithm to explore the possible paths connected to the second node. The Adjacencies database contains way id, from node id and to node id; where the way id is the way this node currently resides in, and from node id and to node id are adjacent nodes. When doing the searching, we can find all the nodes connected to the first node by querying the Adjacencies database, and applying some restrictions to eliminate impossible nodes, for example, you do not have to search the passed nodes. The algorithm will terminate when the second node is reached. Due to the traits of road map, BFS applied here is not very time consuming, as it does in the graph searching. Because each node in OSM has a limited number of connecting points (e.g. crossing node usually has four neighbor nodes), the running time is O(V ), not O(V +E) if there are V nodes and E edges in the road map. III. RESULTS In this section, we present the results of our experiments. The original sampling rate of the GPS was 1 Hz. The path constructed from it is considered as the ground truth when evaluating the path construction accuracy. A. Path Construction Evaluation In order to find the relationship between the sampling rate and the path construction accuracy, we need to collect trace data with different sampling rates. Fig. 4 shows the evaluation process.

5 (a) Two nodes with same way ID (b) Connect two nodes by adding points between them along the road (c) Two nodes with different way IDs (d) Explore path by querying database to find possible path Fig. 3: Path Construction. After map matching, we can construct the path by adding OSM nodes since we cannot connect them directly. If two matched nodes with the same way ID (a), we can add OSM nodes fetched from database which are between them along the road (b). If two matched nodes with different way IDs (c), we would do a BFS search based on the Adjacency database from the first node. Once the second node is reached, the path is found (d) and the algorithm will terminate. Map data c 2013 Google where d is the length erroneously subtracted, d + is the length erroneously added, and d 0 is the length of the route constructed from the original GPS trace (ground truth). After running our path construction algorithm on all 123 traces that we have collected, and comparing them with the path constructed from the original trace, we quantified how the mismatched fraction of route changed when constructing a route from GPS traces with different sample periods. Fig. 7 shows the average mismatch fractions of all traces with different sampling rates. B. Energy Consumption Analysis Fig. 4: Evaluation procedure. The sampling rate will gradually decrease from 1 Hz to Hz. The path constructed from the original GPS sampling rate will be regarded as the ground truth. Along with the increase of the sampling interval, we will compare the constructed paths with the ground truth, and find out the relationship between the mismatch fraction and sampling rate. The data we collected are time stamped latitude/longitude coordinates. To relieve the data collection workload and reduce repetitive work, we can get different sampling rate trace data by simulating the sampling rate. We simply select samples with specific time intervals and remove other samples between them. The sampling interval is gradually increased to 10, 20, 30, 40, 50, 60, 90, 120, 240, 300 seconds. Initially, we ran our map matching algorithm on all of the traces, we got the corresponding map data provided by OSM and filtered out the unqualified data points. A trace example is shown in Fig. 5(a), which is miles having 327 data points. After the map matching, we have 66 nodes for this trace from OSM, as shown in Fig. 5(b). To evaluate the path construction accuracy, we estimate it by the percentage of mismatched fraction of route using equation 1. Fig. 6 explains how we estimate the mismatched fraction. Mismatch = (d + d + )/d 0 (1) After conducting experiments on the GPS traces, we have the tradeoff between GPS sampling rate and route accuracy. However, in order to find the tradeoff between energy consumption and route accuracy, we need to analyze the energy consumption while constructing a path from different sampling rates. Our evaluation does not exactly match GPS duty cycling that can be implemented on software on smartphones. Since it takes time for GPS to get a position fix after acquiring satellite signals [16], [17], we need to take GPS Time To First Fix (TTFF) into account when we simulate reducing the sampling rate. TTFF is a measure of the time required for a GPS receiver to acquire satellite signals and navigation data. However, GPS TTFF is not a constant. It depends on, for example, the weather conditions, the number of visible satellites, signal strength, and driving speed. In order to have a general approximation of TTFF, we test it by activating GPS and timing its TTFF in real driving circumstances under different weather conditions and road conditions. According to National Instrument s GPS Receiver Testing manual [16], TTFF is also related to the status of GPS receiver (cold start, warm start and hot start). We would like the GPS receiver to perform hot start on each cycle, so the duty cycling period is 10 and 20 seconds. Under normal driving conditions, we tested how long it took to get a position fix each time after activating the GPS receiver on average. We

6 Fig. 7: Mismatch fraction vs. Sampling rate. This is the average result of all 123 traces. From the table, we can see that the loss of route accuracy stabilizes around 23% along with the increase of the sampling period between 200s and 300s. have collected 2150 samples. The distribution of GPS TTFF is as shown in Fig. 8. (a) Before Map Matching Fig. 8: Probability distribution of GPS TTFF. The GPS TTFF varies under different conditions, the cumulative probability that the TTFF is less than 20 seconds is higher than 94%. (b) After Map Matching Fig. 5: Map Matching Example. There are a total of 327 nodes in the original trace, shown in (a). After map matching, only 66 OSM nodes are matched and extracted. Map data c 2013 Google We can see from the figure, even though TTFF is not a fixed value, the probability that the GPS receiver takes less than 20 seconds to get a position fix is higher than 94%, which satisfies the data provided by the GPS Receiver Testing manual [16]. The mean GPS TTFF from our experiments is seconds. We suppose that the most energy inefficiency occurs when the GPS TTFF is 20 seconds (worst case), because, in order to obtain one single position fix, the receiver stays on for 20 seconds, which means that it has to decrease the time that the GPS receiver is off during the duty cycling to maintain relatively high accuracy. The average case for battery consumption is when we consider GPS TTFF as the mean value seconds (general case). To evaluate the tradeoff between route accuracy and energy consumption when duty cycling GPS, we first define energy consumption, which is shown as equation 3. Ttotal = Tof f + Ton Pduty cycling = Ton Tof f Psampling + Pidle Ttotal Ttotal where Fig. 6: Error measurement. Compared to the ground truth, the matched route has either erroneously added some fractions, or erroneously subtracted some fractions. If no fraction added or subtracted, then it is a perfect match. Ttotal : time interval between GPS samples Tof f : time that GPS is off during each duty cycling Ton : time that GPS is on during each duty cycling (2) (3)

7 P duty cycling : power consumption when duty cycling GPS P sampling : power consumption when sampling GPS P idle : power consumption when device is idle. Our measurement of power consumption of a GPS receiver was performed on a Samsung Galaxy Player 4.0. It is an mp3 player with a GPS component. We used an oscilloscope to measure the current when the GPS component is working. Because Android devices cannot boot without the battery, we left the battery connected to the device and measured continuous power transferred from the battery to the device. To avoid interference from battery charging circuitry, we took the measurements with no external charger connected to the device. We placed a 0.5 ohm resistor in series with ground and measured the voltage drop on the resistor from which we can calculate the current in the circuit. The power consumption we measured was under a hot start condition, as mentioned above. The device ran a background service to get locations by activating its GPS receiver periodically. During the idling, the current is 8 ma and the power consumption is 29.6 mw, while the current burst when activating the GPS receiver is as high as 260 ma. During the GPS sensing, the current surged to 88 ma and the power consumption is mw. Thus, for a smartphone with a battery capacity of 1200 mah, continuous GPS sampling can drain the battery in 13.6 hours, assuming that the smartphone is not used for anything else. Fig. 10: Average case of tradeoff between route accuracy and energy consumption when duty cycling the GPS receiver. The GPS TTFF we measured had mean of seconds. We used this value to analyze the tradeoff as the average case. T (s) T on(s) T off (s) Mismatched Fraction Avg Power Consumption (mw) Worst Avg Worst Avg Worst Avg Worst Avg TABLE I: Route accuracy and energy consumption in the worst case and the average case. When increasing the sampling period from 20 seconds to 300 seconds, the average energy consumption is decreasing. When the the sampling period goes beyond 120 seconds, it can save at lease 78% energy without sacrificing much route accuracy; it stabilizes around 23%. Fig. 9: Worst case of tradeoff between route accuracy and energy consumption when duty cycling the GPS Receiver. Our experiments show that TTFF has a 94% probability that it does not exceed 20 seconds. Thus, we used 20 seconds to analyze the tradeoff as the worst case. Based on the energy consumption definition and TTFF condition, we measured energy consumption for the worst case and general case, respectively. The worst case means the longest time to get the first fix (TTFF) and the average case means the mean time to get the first fix (TTFF). Table I shows the results. Fig. 9 and Fig. 10 show the changes of mismatch fraction and energy consumption along with different sampling rates for the worse case and average case. From the figures, we can see that the mismatch fraction of the path is increasing while the power consumption is decreasing. When increasing the GPS sampling period beyond 120 seconds, the power consumption remains almost the same; moreover, the mismatch fraction tends to be stable and is not higher than 25%. In other words, we can achieve satisfactory path construction accuracy while saving as much as 78% of the energy. Thus, we can prolong the battery life of our smartphones, and supply for trajectory-based services at the same time. IV. A. Limitations of Map Matching FURTHER CONSIDERATIONS Our approach provides a map matching algorithm with high accuracy. However, it highly depends on the map data provided by OpenStreetMap, which is uploaded by users around the world. The density of map data provided by OSM varies from area to area. In certain areas, the density of nodes in the database is inadequate due to infrequency of GPS trace uploading. For example, in addition to our collected dataset in Central Jersey, we tested a trace we found online in Beijing, China, which contains 1,270 nodes. After map matching, the route is represented by only 15 OSM nodes. For this area, the density of map data is inadequate compared to the complexity and density of the road network. Interestingly, we found that mismatched fractions of a route do not increase steadily as the sampling rate decreases. This situation is caused by the randomness when removing samples in the GPS trace and the characteristic that our algorithm tends to find the shortest path. In the experiment, we found that some paths constructed via a lower sampling rate is more accurate than those constructed via a higher sampling rate. Fig. 12 illustrates how it happens. As we can see, though the 30 seconds sampled trace has denser nodes, it does not contain the key node that is on the curve. When it comes to explore the path between the two nodes, our algorithm tends to find the shortest path between them. Therefore, it constructs the wrong path. To mitigate this situation, we could use some sensors to

8 (a) 30sec sampled trace (b) path constructed from (a) (c) 50s sampled trace (d) path constructed from (c) Fig. 11: Path Construction Error. This is the map matching of Figure 5(a). Due to the key points missing in the 30 seconds sampled trace, the path constructed by it mismatched a fraction of the route. But the path constructed via 50 seconds sampled trace is 100% accurate, it persists in identifying the key points in the curvy road. Map data c 2013 Google. help detect turns when driving. By comparing turning points and curvature of the roads, it is possible to avoid missing key nodes and improve route accuracy [18]. Fig. 12: A route sampled every 30 second compared to every 50 second. Although the 30 second sampling rate trace contains more nodes, it does not contain the key point which resides on the curve. Due to the shortest path characteristic of our algorithm, this will lead to the wrong path construction. We observed that the maximum mismatched fraction happens from 20 seconds to 40 seconds. By visualizing the trace on the map as shown in Fig. 11, we know the mismatched fraction of route usually happens on the local part of the route. For highways, it does not change much as the sampling period increases. Take a residential road for example, the speed limit is 25 mph, which means a driver could travel 0.2 miles within 30 seconds on it. Considering the high road density in local area, 0.2 miles between two nodes means that there could be more than two streets or roads connecting them. Because more than one path are legal, the chance of choosing a wrong path by our path construction algorithm thus increases. However, on a freeway, even though the driving speed is 65 mph, which means a driver could travel 0.5 miles within 30 seconds, most likely there is only one legal path between two nodes due to much less density and complexity of the roads. To improve route accuracy on local roads, instead of using fixed duty cycling, we can use adaptive duty cycling based on the density of the road network. We should take the type of road into account, and quantify the factors that affect route accuracy. B. Tradeoffs in Scheduling GPS Duty Cycling From our results, we know that if we can bear the loss of accuracy on local roads, fixed duty cycling is a feasible approach for energy efficient trajectory-based services if the duty cycling is well scheduled. To schedule the duty cycling, we need to establish some criteria based on the specification that a certain app wants to achieve. We can analyze the trade-off result and make a decision by two decision-making methods: single criterion choice and pros & cons evaluation. 1) single criterion choice: For this method, only one criterion dominates. For trajectory-based services, the criterion for route accuracy usually dominates over energy consumption. So, how to schedule a duty cycle is mainly based on the requirement of route accuracy. We can determine the most efficient duty cycling period based on the maximum loss of accuracy an app can bear. 2) multiple criteria evaluation: We need to consider all the criteria, quantify those criteria by inventing a scoring system and make a decision based on this score. An example is illustrated as follows. An app with trajectory-based service expects to consume less than 250 mw power on GPS sampling. And the maximum mismatched fraction of route it can bear is 0.2. Both of route accuracy and energy consumption are equally weighted to the app. A scoring system can be invented as given by Equations 4, 5, and 6 in the Results section. After calculating the score based on this, a sampling period with the lowest score is the most efficient under those criteria. Where: P unified = P duty-cycling P max (4) F unified = F duty-cycling F max (5) S = P unified W power + F unified W accuracy (6) P max : max power consumption on GPS the app can bear. P unified : power consumption on the scale of P max. P duty-cycling : power consumption when duty cycling GPS. F max : max mismatched fraction of route the app can bear. F unified : mismatched fraction of route on the scale of F max.

9 Sampling Period(sec) P unified W power F unified W accuracy Score TABLE II: Scores of different sampling periods. According to the weight of each criterion and the formula give above, the score can be quantified, and the one with lowest score is the most efficient one. F duty-cycling : mismatched fraction of route when duty cycling GPS. W power : weight of power consumption for the app. W accuracy : weight of route accuracy for the app. S: score under all criteria. Applying the scoring system on the previous results, we have scored each sampling period as in Table II. The sampling period with the lowest score is 20 seconds. Thus for this app, we can schedule the duty cycling of GPS receiver with a sampling period of 20 seconds. V. RELATED WORK Considerable research has been devoted to energy-efficient localization. However, most of the approaches focus on the receiver design and the signal processing algorithms. Sun et al. [19] has written an overview on signal processing techniques used in A-GPS systems. Xiao et al. [20] has written a comprehensive survey on energy efficiency for location based applications. Duty cycling GPS is an obvious solution to prolong the device s battery lifetime at the cost of sacrificing the position accuracy while GPS is off. Raja et al. [4] studied the tradeoff between energy consumption and localization performance for sensor networks. They analyzed the relationship among energy consumption, GPS uncertainty, GPS off-time and speed, and proposed three speed models for duty cycling strategies for maintaining position uncertainty within specified bounds. The idea behind duty cycling is to reduce the position updates frequency by only sampling GPS when the position uncertainty exceeds the accuracy threshold. Rate-adaptive positioning system (RAPS) [3] is proposed based on the observation that GPS is generally less accurate in urban areas, so it suffices to turn on GPS only as often as necessary to achieve this accuracy. RAPS achieves much of its energy saving by avoiding GPS activation when it is unavailable. However, when it comes to driving on the roads, that GPS service is always available and necessary, which in turn does not save much energy. Similarly, relying on the idea that location applications do not always need the highest available accuracy, A-loc [21] proposes to dynamically trade off location accuracy and energy consumption. LEAP [7] explains why it is hard to realize duty cycling GPS receiver for trajectory-based services. In order to be energy efficient, LEAP shifts delay tolerant position calculations to the cloud. It also introduces a mechanism for fast re-acquisition based on previous tracking results. Therefore, LEAP receives good energy-efficiency for trajectory-based services. Generally, it is only necessary to update the location via GPS sampling when the object is moving. Farrell et al. [22] and You et al. [23] take into account constant parameters such as positioning accuracy, delay and target speed, to detect the state of the target. EnTracked [5] uses an accelerometer to detect movement. By detecting movement and estimating speed, the paper proposes an error model which is used to compare the defined error limit to calculate the time limit for the next GPS position update. The main limitation of EnTrack is that the movement detection algorithm is not accurate enough and only suitable for pedestrians with a speed less than 10m/s, which could be problematic when it comes to real-world driving scenario. EnTrackedT [24] is an extension of the EnTracked system. EnTrackedT uses adaptive duty cycling strategies for the accelerometer and compass sensors, which make the system more efficient. And a trajectory simplification algorithm is designed for energy-efficient trajectory update protocols. Bhattacharya et al. [25] extended and further analyzed the sensor management strategies of the EnTrackedT system. Specifically, they proposed the concept of situational bounding to improve and parameterize the robustness of sensor management strategies for trajectory tracking. As periodic or adaptive duty-cycling of GPS may not achieve significant energy savings under all conditions, therefore some works have explored schemes that are based on the spatiotemporal consistency in user mobility, or the large population statistics on routes in an area. Paek et al. [2] proposes an approach, CAPS, to save energy by using a cell-id sequence matching technique to estimate current position. The paper reveals a fact that people often take similar routes in daily life. CAPS only turns on GPS to opportunistically learn and build the history of routes when a user explores new routes. EnLoc [26] also explored how to make use of the spatiotemporal history in a user mobility profile. It uses an offline dynamic program to find the optimal localization accuracy for a given energy budget and decides which one or combination of GPS/WiFi/GSM localization methods to use. SensLoc [27] efficiently provides contextual information about places and paths, instead of simply tracking a user s raw coordinates. The place detector recognizes the place by a scanning radio beacon and consulting a database. Path tracker is only enabled when the place detector senses a place departure. The system saves energy by context detection, only enabling GPS on these paths. All of the above works introduce innovative approaches for energy-efficient localization. Naturally, they include trade-offs between positioning accuracy and adaptive duty cycling GPS. Approaches include, for example, using low power sensors to aid GPS sampling or compensating for GPS sampling insufficiency by using history location information. However, the papers generally argue that using point fixed duty cycling GPS receiver is not a good choice for trajectory-based services. In this paper, we show that with the knowledge of road information, we are still able to construct the route from a GPS trace even when the GPS sampling rate is inadequate, meanwhile we can save up to 78% energy consumption.

10 VI. CONCLUSIONS Duty cycling GPS achieves energy efficiency by sacrificing route accuracy. As the sampling period increases, both energy consumption and route accuracy generally decrease. When the sampling period exceeds 120 seconds, the mismatched fraction of route tends to be stable around 25%. By visualizing routes on the map, we can make a robust assumption that this value is proportional to the composition of the route. When reducing the sampling rate, the mismatched fraction of the route is more likely to occur on local roads rather than on highways. Thus, duty cycling mainly saves energy from the loss of accuracy on the local part of the routes. If a mobile app can bear the loss of accuracy on local roads, fixed duty cycling GPS is a robust approach for energy-efficient trajectory-based services. ACKNOWLEDGMENTS This material is based upon work supported by the National Science Foundation under Grant Number Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. REFERENCES [1] Y. Zheng, Trajectory data mining: an overview, ACM Transactions on Intelligent Systems and Technology (TIST), vol. 6, no. 3, p. 29, [2] J. Paek, K.-H. Kim, J. P. Singh, and R. Govindan, Energy-efficient positioning for smartphones using cell-id sequence matching, in Proceedings of the 9th international conference on Mobile systems, applications, and services. ACM, 2011, pp [3] J. Paek, J. Kim, and R. Govindan, Energy-efficient rate-adaptive gps-based positioning for smartphones, in Proceedings of the 8th international conference on Mobile systems, applications, and services. ACM, 2010, pp [4] R. Jurdak, P. Corke, A. Cotillon, D. Dharman, C. Crossman, and G. Salagnac, Energy-efficient localization: Gps duty cycling with radio ranging, ACM Transactions on Sensor Networks (TOSN), vol. 9, no. 2, p. 23, [5] M. B. Kjærgaard, J. Langdal, T. Godsk, and T. Toftkjær, Entracked: energy-efficient robust position tracking for mobile devices, in Proceedings of the 7th international conference on Mobile systems, applications, and services. ACM, 2009, pp [6] Z. Zhuang, K.-H. Kim, and J. P. Singh, Improving energy efficiency of location sensing on smartphones, in Proceedings of the 8th international conference on Mobile systems, applications, and services. ACM, 2010, pp [7] H. S. Ramos, T. Zhang, J. Liu, N. B. Priyantha, and A. Kansal, Leap: a low energy assisted gps for trajectory-based services, in Proceedings of the 13th international conference on Ubiquitous computing. ACM, 2011, pp [8] J. Liu, B. Priyantha, T. Hart, H. S. Ramos, A. A. Loureiro, and Q. Wang, Energy efficient gps sensing with cloud offloading, in Proceedings of the 10th ACM Conference on Embedded Network Sensor Systems. ACM, 2012, pp [9] OpenStreetMap, 2014, [10] S. Brakatsoulas, D. Pfoser, R. Salas, and C. Wenk, On map-matching vehicle tracking data, in Proceedings of the 31st international conference on Very large data bases. VLDB Endowment, 2005, pp [11] J. S. Greenfeld, Matching gps observations to locations on a digital map, in Transportation Research Board 81st Annual Meeting, [12] Y. Lou, C. Zhang, Y. Zheng, X. Xie, W. Wang, and Y. Huang, Mapmatching for low-sampling-rate gps trajectories, in Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. ACM, 2009, pp [13] P. Newson and J. Krumm, Hidden markov map matching through noise and sparseness, in Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. ACM, 2009, pp [14] M. Attia, A. Moussa, and N. El-Sheimy, Map aided pedestrian dead reckoning using buildings information for indoor navigation applications, Positioning, vol. 2013, [15] Y. Chen and J. Krumm, Probabilistic modeling of traffic lanes from gps traces, in Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems. ACM, 2010, pp [16] N. Instruments, Gps receiver testing, Mar [17] G. Navstar, User equipment introduction, Department of Defense Document MZ10298, vol. 1, [18] X. Gao, B. Firner, S. Sugrim, V. Kaiser-Pendergrast, Y. Yang, and J. Lindqvist, Elastic pathing: Your speed is enough to track you, in Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ser. UbiComp 14. New York, NY, USA: ACM, 2014, pp [Online]. Available: [19] G. Sun, J. Chen, W. Guo, and K. R. Liu, Signal processing techniques in network-aided positioning: a survey of state-of-the-art positioning designs, Signal Processing Magazine, IEEE, vol. 22, no. 4, pp , [20] X. Ma, Y. Cui, and I. Stojmenovic, Energy efficiency on location based applications in mobile cloud computing: a survey, Procedia Computer Science, vol. 10, pp , [21] K. Lin, A. Kansal, D. Lymberopoulos, and F. Zhao, Energy-accuracy trade-off for continuous mobile device location, in Proceedings of the 8th international conference on Mobile systems, applications, and services. ACM, 2010, pp [22] T. Farrell, R. Cheng, and K. Rothermel, Energy-efficient monitoring of mobile objects with uncertainty-aware tolerances, in Database Engineering and Applications Symposium, IDEAS th International. IEEE, 2007, pp [23] C.-W. You, P. Huang, H.-h. Chu, Y.-C. Chen, J.-R. Chiang, and S.- Y. Lau, Impact of sensor-enhanced mobility prediction on the design of energy-efficient localization, Ad Hoc Networks, vol. 6, no. 8, pp , [24] M. B. Kjærgaard, S. Bhattacharya, H. Blunck, and P. Nurmi, Energyefficient trajectory tracking for mobile devices, in Proceedings of the 9th international conference on Mobile systems, applications, and services. ACM, 2011, pp [25] S. Bhattacharya, H. Blunck, M. B. Kjærgaard, and P. Nurmi, Robust and energy-efficient trajectory tracking for mobile devices, Mobile Computing, IEEE Transactions on, vol. 14, no. 2, pp , [26] I. Constandache, S. Gaonkar, M. Sayler, R. R. Choudhury, and L. Cox, Enloc: Energy-efficient localization for mobile phones, in INFOCOM 2009, IEEE. IEEE, 2009, pp [27] D. H. Kim, Y. Kim, D. Estrin, and M. B. Srivastava, Sensloc: sensing everyday places and paths using less energy, in Proceedings of the 8th ACM Conference on Embedded Networked Sensor Systems. ACM, 2010, pp