How Long to Wait? Predicting Bus Arrival Time With Mobile Phone Based Participatory Sensing

Editor's Notes

• #2: Hi everyone, I’m today’s first speaker Caroline The paper I want to present is How Long to Wait? Predicting Bus Arrival Time With Mobile Phone Based Participatory Sensing. It comes from IEEE Transactions on Mobile Computing at June 2014
• #3: This is the outline: First is introduction, and the related work. And then I will explain the design and challenges of this system and describe the technical solutions. And see the implement and evaluation at Singapore and London, finally is the conclusion of this paper.
• #4: Ok Let’s begin at the introduction and related works
• #5: Public transport, especially the bus transport, has been well developed in many parts of the world. The bus transport services reduce the private car usage and fuel consumption, and alleviate traffic congestion. As one of the most well-known means of public transport, The bus system serves over 3.3 million bus rides every day with around 5 million residents in Singapore.
• #6: When traveling with buses, the travelers usually want to know the accurate arrival time of the bus. Excessively long waiting time at bus stops may drive away the anxious travelers and make them reluctant to take buses. In order to fix this problem, there are some kinds of solution:
• #7: Most bus operating companies have been providing their timetables on the web freely available for the travelers. However, the bus timetables only provide very limited information (like operating hours, time intervals) which are typically not timely updated according to instant traffic conditions.
• #8: Other than those official timetables, there are many public services like Google Maps are provided for travelers. Although such services offer useful information, they are far from satisfactory to the bus travelers. For example, the schedule of a bus may be delayed due to many unpredictable factors (like traffic conditions). The public services might not know the information in time.
• #9: The accurate arrival time of next bus will allow travelers to take alternative transport choices, and improve their experience. So towards this aim, many commercial bus information providers offer the real-time bus arrival time to the public. But providing such services, usually requires the cooperation of the bus operating companies And need to install special location tracking devices like GPS on the buses Which will come with a lot of cost.
• #10: There are some related work about this issue, one is Phone-based transit tracking. And there are some kind of system has been proposed, the difference is our system use cell tower sequence information. And no explicit location services are invoked so as to reduce the overhead.
• #11: The drawbacks of GPS include: Although many GPS-enabled mobile phones are available on the market, a good number of mobile phones are still shipped without GPS modules [26]. Besides, GPS module consumes substantial amount of energy, it will significantly reduce the lifetime of mobile phones [26]. The mobile phones in vehicles may perform poorly when they are placed without line-of-sight paths to GPS satellites [10].
• #12: In this system design, they avoid the need for accurate bus localization. Since the buses travel on certain bus routes, the knowledge of the current position and the average velocity of the bus are enough to predict its arrival time at a bus stop.
• #13: For instance, say the bus is currently at bus stop 1, and a user wants to know its arrival time at bus stop 6. Accurate prediction of the arrival time requires the distance between bus stop 1 and 6 along the 1D bus route and the average velocity of the bus. In general, the physical positions of the bus and the bus route on the 2D maps are not strictly necessary.
• #14: In this paper, instead of pursuing the accurate 2D physical locations they logically map the bus routes to a space featured by sequences of nearby cellular towers. They classify and track the bus statuses in such a logical space so as to predict the bus arrival time. So the bus route will be map to the cell tower sequence {1, 2, 4, 7, 8, 4, 5, 9, 6}
• #15: So in this paper, they present a novel bus arrival time prediction system based on crowd-participatory sensing. The system is independent of the bus operating companies or other third-party service providers. Allowing easy and inexpensive adoption of the proposed approach over other region.
• #16: Second, based on the mobile phones, the system doesn’t need special hardware or extra vehicle devices which can substantially reduces the cost and be more energy-friendly. And this system can automatically detecting environments and generating bus route related reports Means it doesn’t require the human inputs from the participants, which can facilitates the involvement of participatory parties.
• #17: Though the idea is intuitive, the design of such a system in practice entails substantial challenges. In this section, I will describe the major components of the system design. Explain the challenges and present several techniques to cope with them.
• #18: This is the architecture of bus prediction system. There are 3 major components. Sharing users, Querying users and backend server I will introduce them step by step
• #19: The sharing user contributes the mobile phone sensing information to the system. Since the sharing user may travel with different means of transport, the mobile phone needs to first detect whether the current user is on a bus or not. After a sharing user gets on a bus, the data collection module starts to collect a sequence of nearby cell tower IDs. The collected data is transmitted to the server via cellular networks. Ideally, the mobile phone of the sharing user will automatically performs the data collection and transmission without the manual input from the sharing user.
• #20: And querying users query the bus arrival time by sending the request to the backend server. The querying user needs to indicates the interest bus route and bus stop to receive the predicted bus arrival time.
• #21: We shift most of the computation burden to the backend server. There are two stages are involved in this component. In order to bootstrap the system, we need to survey the corresponding bus routes in the offline pre-processing stage. And the server will construct a basic database that associates particular bus routes to cell tower sequence signatures. Since we do not require the absolute physical location reference we can mainly war-drive the bus routes and record the sequences of observed cell tower IDs This can significantly reduce the initial construction overhead.
• #22: In the experiments they found out that even if a passenger travels by the same place, the connected cell tower might be different from time to time due to varying cell tower signal strength. So instead of using the associated cell tower, they record a set of cell tower IDs that the mobile phone can detect.
• #23: To validate such a point, they do an initial experiment. They measure the cell tower coverage at two positions A and B within the university campus, which are about 300 meters. The right picture report the cell tower that the mobile phone can detect, as well as their average signal strength and connection time at A and B, respectively The position A and position B are both covered by 6 cell towers with divergent signal strength. At position A the mobile phone is connected to the cell tower 5031 over 99% of the time, while its signal strength remains consistently the strongest during the 10-hour measurement. And the mobile phone at position B observes two cell towers with comparable signal strength.
• #24: They find that the mobile phone is more likely to connect to the cell tower with stronger signal strength, and also may connect to the cell tower with the second strongest signal strength. Nevertheless, during 7-week experiments, they consistently observe that mobile phones almost always connect to the top-3 strongest cell towers. Therefore, in practice we choose the set of the top-3 strongest cell towers as the signature for route segments.
• #25: This figure shows the cell tower sequence collected on their campus bus traveling from school to a rapid train station off the campus. The whole route of the bus is divided into several sub-route segments according to the change of the top-3 cell tower set. They are marked in red and black in this figure. For example, the mobile phone initially connects to cell tower 5031 in the first sub-route and the top-3 cell tower set is {5031, 5092, 11141}. Later the mobile phone is handed over to cell tower 5032. Such a sequence of cell tower ID sets identifies a bus route in the database. By war-driving along different bus routes, it can easily construct a database of cell tower sequences associated to particular bus routes.
• #26: When the database is ready, the backend server can processes the cell tower sequences from sharing users in the online processing stage. Receiving the uploaded information, the backend server first classifies the uploaded bus routes primarily with the reported cell tower sequence information. The bus arrival time on various bus stops is then derived based on the current bus route statuses.  
• #27: Implementing such a participatory sensing based system, however, entails substantial challenges. First is bus detection , since the sharing users may travel with diverse means of transport we need to first let their mobile phones accurately detect whether or not the current user is on a bus. Without accurate bus detection, mobile phones may collect irrelevant information, leading to unnecessary energy consumption or even inaccuracy in prediction results.
• #28: Nowadays, IC cards are commonly used for paying transit fees in many areas On a public bus in Singapore, several card readers are deployed for collecting the fees. When a passenger taps the transit card on the reader, the reader will send a short beep audio response to indicate the successful payment. In our system, we choose to let the mobile phone detect the beep audio response of the card reader, since such distinct beeps are not widely used in other means of transportation such as non-public buses and taxis.
• #29: The bottom figure plots the raw audio signal in the time domain, where the IC card reader starts beeping about from 11000th sample and lasts to 18000th sample. After we convert the time domain signal to the frequency domain through 512pt Fast Fourier Transform (FFT), we observe clear peaks at 1kHz and 3kHz frequency bands. For comparison, figure (a) is no beep signal. We find no peaks at 1kHz and 3kHz frequency bands.
• #30: With the knowledge of the frequency range of the beep signal sent by the IC card reader we can lower down the audio sampling rate of the mobile phone to 8kHz (8000 samples/s) which is sufficient to capture the beep signals with maximum frequency of 3kHz. And in practice 128pt FFT suffices to detect the IC card reader on the bus with tractable computation complexity on mobile phones.
• #31: We use the standard sliding window averaging technique with window size w = 32 samples to filter out the noises. And use an threshold ε of three standard deviation (i.e., 99.7% confidence level of noise) to detect beep signals. In this figure, when the IC card reader starts beeping, the signal strengths in both 1kHz and 3kHz frequency bands jump significantly and therefore can be detected.
• #32: The most possible false positives are from MRT system, because the IC card systems are also deployed in rapid train stations in the entrances. We want to use the accelerometer sensor on the mobile phone to reduce such false detection. The rapid trains are moving at relatively stable speeds with few abrupt stops or sharp turns.
• #33: We collect the accelerometer data at a moderate sampling rate of 20Hz. The raw accelerometer readings are first made orientation-independen. This figure plots the accelerometer readings on a rapid train and a public transit bus. The bottom figure shows the variance of the accelerometer. The variance on the bus is significantly larger than that on the train. Therefore, we can distinguish the buses from the trains using the variance of accelerometer readings by setting a proper threshold. After experiment, they select an threshold 0.03 to balance the false negative and false positive.
• #34: When a sharing user gets on the bus, the mobile phone samples a sequence of cell tower IDs and reports the information to the backend server. The backend server aggregates the inputs from massive mobile phones and classifies the inputs into different bus routes. We need to carefully classify the bus route information from the mixed reports of participatory users. Without users’ manual indication, such automatic classification is more difficult.
• #35: We match the received cell tower sequences to those signature sequences store in the database. This picture shows an example where a sharing passenger gets on the bus at location A. The backend server will receive a cell tower sequence of {7, 8, 4, 5} when the sharing user reaches location B. The cell tower sequence of the bus route stored in the database is 1, 2, 4, 7, 8, 4, 5, 9, 6, then the sequence 7, 8, 4, 5 matches this bus route as a sub-segment.
• #36: In practical scenarios, the sequence matching problem becomes more complicated due to the varying cell tower signal strength. In the matching process, we design a top-k cell tower sequence matching scheme by modifying the Smith-Waterman algorithm [28]. In a top-k set S = {c1, c2, . . . , ck} ordered by signal strength which si ≥ sj, 1 ≤ i ≤ j ≤ k, where ci and si denote cell tower I and its signal strength. The uploaded cell tower sequence from a sharing user is denoted as Sequpload = {u1u2 . . . um)} where m is the sequence length. We also denote a cell tower set sequence in database as Seqdatabase = {S1,S2 . . . Sn} where n is the set sequence length.
• #37: If ui = cw ∈ Sj, ui and Sj are considered matching with each other, and mismatching otherwise. We assign a score f (sw) for a match, where f (sw) is a positive non-decreasing scoring function and w is the rank of signal strength. In practice, f (sw) = 0.5w−1 as the scoring function according to the signal strength order in the set. The penalty cost for mismatches is set to be −0.5 which balances the robustness and accuracy in practice.
• #38: Table 2 shows a cell tower set sequence matching instance. In the example, the uploaded cell tower sequence is Sequpload = {1, 8, 10, 15, 16 } And the cell tower ID set is shown in the first three rows sorted in decreasing order of the associated cell tower signal strength.
• #39: One problem of the cell tower sequence matching is some bus routes may overlap with each other. So they survey 50 bus routes in Singapore and measure their overlapped road segments using Google Maps. This picture shows the distribution of the lengths of overlapped road segments, which suggests that over 90% of the overlapped route segments are shorter than 1400 meters, and over 80% are less than 1000 meters. They set the threshold of cell tower sequence length to 7.
• #40: This picture shows the cell tower sequence matching accuracy in classifying the bus routes We vary the length cell tower sequence from 2 to 9. And find that the matching accuracy is low when the cell tower sequence length is small (e.g, <4) because of the problem of route overlap. When the cell tower sequence length reaches 6, the accuracy increases substantially to around 90%.
• #41: In many practical scenarios, the length of the cell tower sequence obtained by a single sharing user, may be insufficient for accurate bus route classification. So we can concatenate several cell tower sequences of different sharing users on the same bus to form a longer cell tower sequence. In this example, both cell tower sequences of sharing user A and B are short, But by concatenating the two cell tower sequences the backend server may obtain an long cell tower sequence which can be used for more accurate bus classification.
• #42: A simple way of concatenating the cell tower sequences is to let the mobile phones of sharing user locally communicate with each other (e.g., over Bluetooth) [20]. But this approach, mandates location exposure among sharing user and might raise privacy concerns. So we shift such a job to the backend server.
• #43: Recall that the mobile phone needs to collect audio signals for bus detection. Here, we reuse this information to detect whether the sharing passengers are on the same bus. This picture shows an example of the audio signals captured by three different mobile phones on the same bus. We depict the raw audio signals in Fig. 12(a), and corresponding frequency domain signals in Fig. 12(b)–(d). We can see that in the frequency of domain the signals are highly cross-correlated and thus can be used to determine whether the phones are on the same bus. Such beep interval information is reported along with the cell tower sequences to the backend server. Receiving the uploaded sensing data from sharing user, the backend server detects and groups the sharing user on the same bus by comparing both cell tower sequences and the time intervals of the beep signals.
• #44: After the cell tower sequence matching, the backend server classifies the uploaded information according to different bus routes. When receiving the request from querying users the backend server looks up the latest bus route status, and calculates the arrival time at the particular bus stop.
• #45: This figure illustrates the calculation of bus arrival time prediction. The server needs to estimate the time to travel from its current location to the queried bus stop. Suppose that the sharing user on the bus is in the coverage of cell tower 2, the backend server estimates its arrival time at the bus stop according to both historical data and the latest bus route status.
• #46: The server first computes the dwelling time of the bus at the current cell (i.e., cell 2 in this example) denoted as t2. The server also computes the traveling time of the bus in the cell that the bus stop is located denoted as tbs. The arrival time of the bus at the queried stop is then estimated as this equation.
• #47: In generality, we denote the dwelling time in cell i as Ti, 1 ≤ i ≤ n, the bus’s current cell number as k, and the queried bus stop’s cell number as q. The server can estimate the arrival time of the bus as this equation The server periodically updates the prediction time according to the latest route report from the sharing users and responds to querying users.
• #48: And now we can move on to the Implement & Evaluation
• #49: They implement the mobile phone applications with the Android platform using Samsung Galaxy S2 and HTC Desire. Both types of mobile phones are equipped with accelerometers and support audio signal sampling from microphones. And they implement the backend server in Java running on the DELL Precision.
• #50: They experiment on both campus shuttle buses and public transport As this picture show, there are 4 shuttle bus routes in campus. The bus route lengths span about from 3.8km to 5.8km with cell tower set sequence lengths varying from 9 to 22. Table 3 gives the details of the bus routes. Table 4 summarizes the shared route segments between each pair of bus routes
• #51: They collect more than 200 beep signals on different public transit buses. And set the audio sampling rate to be 8kHz, and use 128-pt FFT to detect the IC card reader. The distances between the IC card reader and the mobile phones vary from 1 meter to 7 meters. We also consider the scenarios where mobile phones may be held in hand and inside bags. We see that the detection rate is over 95% when mobile phones are close to the IC card reader, even when they are placed in bags. As the distance increases further, the detection accuracy drops substantially.
• #52: Next we evaluate the accelerometer based bus detection method that is used to distinguish the buses from the MRT trains. This figure plots the accuracy in detecting the buses. We find that accelerometer based method can distinguish the buses from the MRT trains with an accuracy of over 90% on average. //The main reason for falsely detecting public buses as MRT trains, is because it happens mostly when the buses are driving along long straight routes late during night time. //The accelerometer readings may be relatively stable and very similar to those on the MRT trains.
• #53: Although 85% of Route A is overlapped with Route C, the bus classification accuracy for Bus A and C are still around 94%. The main reason is that Bus A and C travel in the opposite directions. Since Route D shares a large portion of overlapped road segments with A and C, and buses travel in the same direction on the shared road segments, buses along Route D might be misclassified to A or C.
• #54: The median prediction errors vary about from 40s (Bus B) to 60s (Bus D). The 90th percentiles are about from 75s (Bus B) to 115s (Bus D). The average estimation error increases as the length of bus route increases. We observe that as the bus moves closer to the querying user, the prediction error becomes smaller.
• #55: We experiment with commercial bus system as well. For comparison, we also query the arrival time of public transit buses provided by LTA of Singapore. The public buses are periodically tracked with on-bus localization devices and respond to the queries for the bus information. Both prediction errors of LTA and our system are measured and plot in this Figure. According to the results, the average prediction error of our system is about 80 seconds, while the prediction result of LTA is around 150 seconds. Means our system is more accurate than GPS-based prediction.
• #56: The major computational complexity is attributed to performing Fast Fourier Transform on mobile phones which is O(n log n). Current mobile phones can finish the computation task in real-time. Table 5 illustrates the measured battery lifetime when the mobile phones continuously trigger different sensors. We can see the battery lifetime is substantially reduced when the GPS module in the phone is enabled.
• #57: The computation overhead of backend server is mainly bounded by the bus classification algorithm : The uploaded cell tower sequence length l, cell tower set sequence length k, And number of cell tower set sequences in the database N. The computation complexity of sequence matching using dynamic programming is O(lk) Need to compare with N candidate sequences in database the overall computation complexity is O(lkN). Since in practice both l and n are usually small, the computation complexity increases almost linearly.
• #58: This is another implement in London
• #59: In addition to Singapore, this paper also do trial experiments with London bus system. London Buses Services Limited manages one of the largest bus networks in the world. Oyster cards [4] are widely used for paying the transit fees on London buses.
• #60: As this figure, the experiment is with London bus route 27, 7, 36, 23 and 159. They collect the audio beeps on buses and the cellular signals observed along the bus routes. The bus route details are summarized in Table 6. //The bus route lengths span from 11km to 20km and the cell tower sequence lengths along the bus routes are 48 to 67. //The average bus speed is about 23km/h. The overlapped bus route segments are mainly in the city center.
• #61: Different from Singapore buses, the beep from London buses is a single frequency audio signal of 2.4kHz unique frequency. So they record the beep audio signal from the card readers at a sampling rate of 44.1kHz and extract the frequencies using 512pt FFT. The distances to the nearest card reader to collect beep audio is vary from 1m to 5m. Some of the London buses has second floor. For all the testing positions, we consider the scenarios where the mobile phone may be held in hand or placed inside bags. Table 7 summarizes the average detection accuracy of single beeps. The average detection accuracy (Table 7) is above 80% when the distance is within 4m. The audio detection accuracy decreases as the distance increases.
• #62: In this picture we can see that, there are many overlapped route segments between the 5 bus routes. So we use different lengths of cell tower sequences to perform the bus classification.
• #63: The classification accuracy increases as the cell tower sequence length grows. When the length is longer than 8, the classification accuracy becomes higher than 90%.\ The cell tower sequence collected from the overlapped bus route segments contributes less to those collected from non-overlapped segments.
• #64: Fig b plots the individual performance of bus classification of the 5 experimented bus routes. The route 27 has the shortest overlapped route segments while route 23 has the longest This cause route 27 has highest classification accuracy, and route 23 has lowest classification accuracy But the overall bus classification accuracy is higher than 85% for all bus routes.
• #65: The backend server estimates the bus location with the uploaded cell tower sequences and predicts the bus arrival time based on the cell tower sequence stored in the database. The overall prediction error in this graph is calculated by comparing the predicted and the actual bus arrival time. We can see that the prediction error of route 27 is the lowest and the route 36 is the highest. For the 5 bus routes, the median prediction error varies from 65s (route 27) to 125s (route 36)
• #66: The overall prediction error of bus arrival time in London is higher than that in Singapore mainly due to the following reasons. First, the lengths of the experimental bus routes in London are much longer than in Singapore, which brings more unpredictable factors influencing the bus operation.
• #67: Second, the time duration between two adjacent buses of some bus routes in London is much longer than that in Singapore, which usually results in far away buses from the querying user. Third, the traffic conditions in London are much more complicated than in Singapore. Many unpredictable factors like traffic jam, adaptive traffic lights, pedestrians, etc., may affect the system performance.
• #68: And finally is the conclusion
• #69: In this paper, the author present a crowd-participated bus arrival time prediction system by using inexpensive and widely available cellular signals, which provides cost-efficient solutions to the problem. And being independent of any support from bus operator, the proposed scheme provides a flexible framework for participatory contribution of the community. For a particular city, the only requirement of the system implementation is that there exist a backend server and an IC card based bus system.
• #70: Future work includes how to encourage more participants to bootstrap the system because the number of sharing passengers affects the prediction accuracy in the system. At the initial stage, they may encourage some specific passengers (like the bus drivers) to install the mobile phone clients. The more precise they can predict, the more user will want to participate in.