Application of Bluetooth Technology for Mode-Specific Travel Time Estimation on Arterial Roads: Potentials and Challenges

Bahar Namaki Araghi1*, Rajesh Krishnan2, Harry Lahrmann3 1. PhD Student, Traffic Research Group, Department of Planning and Development, Aalborg University, Fibigerstrde 11, 9220 Aalborg st, Denmark, Tel: +45-9940-8447, E-mail: araghi@plan.aau.dk 2. Centre for Transport Studies, Imperial College London, United Kingdom 3. Associate Prof., Traffic Research Group, Aalborg University, Denmark
1. Introduction and Background Since 2005, Bluetooth Technology (BT) has been used for traffic measurement. The Bluetooth protocol uses an electronic identifier, or tag, in each device called the Machine Access Control (MAC) address. The MAC address serves as a nickname for electronic devices to keep track of who is who during data communications. Vehicles containing detectable Bluetooth devices can be observed by Bluetooth sensors installed on the road side. Bluetooth sensors record MAC addresses of detected devices and the time of detection. Hence, the time stamped MAC addresses recorded by multiple Bluetooth sensors can be used for travel time estimation (Haghani et al., 2012). Increasing number of Bluetooth devices among road users, Bluetooth Sending MAC anonymity of MAC addresses, flexibility of deployment and Sensors Address maintenance of Bluetooth sensors, along with the acceptable accuracy of travel time estimation using Bluetooth Distance 4 Km MAC addresses technology (Haghani et al., 2009; Puckett and Vickich, & Timestamp 2010; Barcelo et al., 2010) have resulted in an emerging sent to ITS interest for further evaluation of the application of this Server technology as a cost-effective method for traffic Bluetooth Device measurement. detected at time t1= 11:10:23 Due to the anonymity of MAC addresses (i.e. no current data Travel Time = t2-t1= 00:05:17 sources link MAC addresses with personal information), Speed = 4/ 05:17= 45.43 Km/h external systems (i.e. Bluetooth sensors) tracking these Bluetooth Device electronic devices (i.e. Bluetooth enabled devices) is more detected at time acceptable to the public compared to Global Positioning t2= 11:15:40 Systems (GPS) or Automated Number Plate Recognition Figure 1. Application of Bluetooth technology for (ANPR). Furthermore, increasing numbers of Bluetooth Travel Time estimation devices not only do not impose any additional cost to the ITS infrastructure but also provides better opportunities for estimating travel time based on MAC address tracking. Larger number of Bluetooth devices increases the penetration rate and correspondingly the sample size of the Bluetooth dataset. It helps to improve the accuracy and reliability of Bluetooth based travel time estimations. Moreover, BT is less sensitive to the weather conditions (i.e. fogs, darkness, etc.) compare to ANPR and GPS which is another positive aspect of this technology and flexibility of its deployment. Although, the Bluetooth has been demonstrated as a promising technology, there remains problems such as overcounting of a single vehicle carrying more than one Bluetooth enabled device (for example buses), low sampling rate and the difficulty of distinguishing between multiple travels modes (e.g. motor vehicles, bicycles and pedestrians). Research to date has primarily focused on the application of BT for travel time estimation on motorways, where there is no need to distinguish between multiple travel modes due to the uniformity of traffic. However, arterial roads and other city streets experience a mix of transport modes (i.e. motor vehicles, bicycles and pedestrians). Therefore, applying Bluetooth technology to estimate mode-specific travel time under heterogeneous traffic conditions remains a challenge. Heterogeneous traffic can affect both accuracy and reliability of travel time estimated using Bluetooth technology. This problem is particularly relevant in cities with a high volume of bicycles traffic. This paper aims to develop a method for estimating mode-specific travel time using MAC address tracking data under uncongested traffic conditions. Since the travel time of different transport modes is different under uncongested conditions (e.g. motor vehicle and bicycles), this paper presents methods to classify the recorded MAC addresses into different groups corresponding to different transport modes.

2. Methodology In order to distinguish bicycles and motor vehicles using Bluetooth data, classification methods can be considered as one potential solution. However, noted that any solution for this problem should be based on both the attributes of the recorded MAC address as input data and the expected output (i.e. mode-specified travel time estimation). For any detected MAC address, the information of timestamp (i.e. detection time), Radio Signal Strength Indicator (RSSI), Class of Device (CoD) are recorded in the feature vector. Although, there are limited numbers of attributes on feature vector to implement, but at the same time, normally the size of Bluetooth datasets is fairly large. In this context, clustering methods are used as the core technique for classification. The above mentioned features would confine the choice of classification technique. The three most applicable clustering mechanisms including Hierarchical clustering, K-means clustering, and two-step clustering techniques are employed. The idea is to classify Bluetooth detection into two groups of motor vehicle and bicycles through these three clustering techniques based on some hypothesis regarding the uncongested traffic situation. In this context, two main hypotheses are determined which aim to test the possibility of classifying Bluetooth recorded MAC address into two groups of bicycles and motor vehicles. 2.1 Hypothesis In order to develop the clustering techniques for both bicycles and motor vehicles travel time estimation using Bluetooth technology, the two following hypotheses are determined. 2.1.1 Hypothesis 1: Difference in travel time First, it is assumed that in uncongested urban arterials the travel time of motor vehicles are statistically significantly lower than the bicycles (in other words motor vehicles are usually faster than the bicycles). In order to test the first hypothesis, the null hypothesis and the alternative hypothesis are defined as follows: Where: Mean travel time in cluster th The indicates that mean travel time in both clusters are similar. The indicates that mean value in both clusters are dissimilar. 2.1.2 Hypothesis 2: Difference in distribution of Bluetooth devices In order to evaluate the first hypothesis, the second hypothesis is adopted. This mainly focused on distribution of different types of Bluetooth devices used by various road users. Due to differences in application of Bluetooth enabled devices (i.e. navigator, phones, computers etc.), it is expected that some classes of devices will only be used by a specific mode. In this case, it is anticipated that navigators and video/audio devices (i.e. headsets and speakers etc.) will be clustered in the group of devices with short travel time (i.e. motor vehicles). As a result it is expected that after clustering of Bluetooth based travel time, detected navigator MAC address should be allocated to the cluster with short travel time. This hypothesis is mainly aimed to test the output of the first hypothesis. As mentioned, the recorded Bluetooth MAC address contains the information about the CoD. The CoD field has a variable format which includes the information of type of device. In order to test this hypothesis, the distribution of CoD will be tested for each cluster. 2.2 Model Validation In order to evaluate the accuracy of different aggregation methods as well as determining the most accurate aggregation method, three accuracy measures including Root Mean Squared Error (RMSE), Mean Percentage Error (MPE) and Mean Absolute Percentage Error (MAPE) are calculated. The MPE, MAPE and RMSE formula are given as follow:
( | ) | Where: The total number of observed motor vehicles and cyclists Real travel time for the ith record time period obtained using camera recording as the ground truth Estimated mean travel time using Bluetooth technology for the ith time period Note: for more information see Araghi (2010).

The travel time obtained by video cameras is used as the ground truth values and the basis of analysis. The estimation of accuracy is then made by comparing Bluetooth-based and camera-based travel time estimations on specified time periods (i.e. 15 minutes). 3. Experiment Setup and Data Collection A section of Sauersvej located in south end of Aalborg is selected as the test-bed. This link is mainly surrounded by Point A green areas on both sides with no access road connection. It has an acceptable distance from residential areas with minimum number of pedestrians plus a flat topography. This link functions as a connector with appropriate numbers of Distance 550m bicycles and motor vehicles (see Figure 2). Two sources of data are used for conducting the evaluation, including BT and camera data. In order to validate the results of the BT, real Point B traffic is recorded through video cameras which used as the ground truth. The Bluetooth sensors implemented in this study are developed Figure 2. Selected link of Sauersvej, Aalborg, by the company BLIP Systems. The sensors known as Denmark BlipTrack consist of Bluetooth detectors sending data via a GPRS connection in real time and providing time stamped of passing Bluetooth devices through a back-end server. The BlipTrack sensors have three directional antennas which could cover a wide range of about 100-300 meters and could provide the directional information of passing Bluetooth devices (see Figure 3). Due to the wide detection zone of the BlipTrack, a single MAC address is likely to be detected more than once while it passes a sensor location. On the one hand, having a wide range of detection events increases the probability of detecting devices with high speed. On the other hand, this also increases the ambiguity and uncertainty about the location of detected devices at the time that they have been tracked by the sensor. Having a number of detection events for each device, raises the question as to which values should be taken to obtain the most accurate estimate. In order to have a direct comparison between the travel time Figure 3. Configuration of BlipTrack antenna and estimated by Bluetooth and camera, in this study multiple the detection zone detections of a unique MAC id are used to measure travel time between point A and B. The results of the travel time are aggregated based upon four simple methods. These methods are the Minimum (BTMin), Maximum (BTMax), Average (BTAvg) and Maximum signal strength (BTMaxRSSI) travel time obtained using all the detections recorded for a unique MAC id. This enables to have an evaluation on the aggregation method as well as the accuracy of the estimations. 4. Data Analysis and Results Results of the analysis show that Bluetooth data have a good overlap with the records of cameras for both motor vehicle and bicycles. This demonstrates the feasibility of BT for representing and estimating the mode-specific travel time. Results of the cluster analysis of the BT show that there is a good potential for implementation of clustering techniques as the method of classifying the transport modes on arterial roads in an uncongested traffic situation. In this context, the results of the clustering technique prove the H0 (i.e. there is statistically significant difference in motor vehicles and bicycles travel time obtained by BT). In other words, it indicates the feasibility of application of BT for mode-specific travel time estimation on urban arterials. Moreover, results of the cluster membership analysis using CoD, prove the second hypothesis regarding the CoD distribution of Bluetooth devices in two clusters. This also validates the accuracy and efficiency of clustering techniques and predetermined hypothesis for mode-specific Bluetooth travel time estimation. In addition, this shows that the CoD can be used as an informative source of data for classifying the traffic modes in mixed-traffic situations. Results of the first and second hypothesis analysis will be presented in the final paper.

5. Conclusion In order to evaluate the feasibility of Bluetooth technology for estimating mode-specific travel time in mixed-traffic situation, different aggregation techniques as well as clustering methods are proposed. Data are collected from an uncongested urban arterial using both Bluetooth sensors and video recording cameras. Final results are presented and evaluated using MPE, MAPE and RMSE as the three accuracy measures. Results of mode-specific travel time estimation using Bluetooth technology in a mixed-traffic prove the capability and feasibility of this method for accurate travel time estimation in an uncongested urban arterial under mixed-traffic conditions. It also shows that aggregation of Bluetooth data based on different estimators will improve the accuracy of travel time estimation. This also emphasises on the importance of finding optimum aggregation technique. For the mode-specific travel time estimation, clustering techniques are proven to be powerful methods for classification and representation of travel time variation between different transportation modes. Although the results of the three techniques are very close, but in most cases the K-means clustering gives the most accurate estimate. However, in terms of automatic optimisation of number of clusters the two-step clustering methodology seems to be more efficient. The accuracy of the clustering techniques also emphasises the validity of the predetermined hypothesis. These show that in an uncongested traffic situation, there is a significant difference between the average travel times of various traffic modes which also can be captured by Bluetooth sensors. Moreover, according to the second hypothesis there is a difference in distribution of the BT devices which can be determined based upon CoD and can be used as a valuable source of information for distinguishing bicycles and motor vehicles. It is worth mentioning that the experiments and conclusions in this study were based on data collected from an uncongested traffic. Special conditions under congested traffic condition or special events (i.e. incidents) were not investigated. However, the clustering methodology can potentially be applied to congested traffic conditions which need further modification and improvement. In this regard, there is a need for further investigation of the clustering techniques as well as quality of mode-specific travel time estimation based upon BT, which will be presented in the final work. References
1. 2. 3. 4. 5. 6. 7. Haghani, A, Yang, S, & Hamedi, M, 2007, 'Cellular Probe Data Evaluation, Case 13 Study: The Baltimore Multimodal Traveler Information System (MMTIS) ', Technical 14 Report, Submitted to Maryland Department of Transportation. Haghani A, Hamedi, A, & Farokhi Sadabadi, K, 2009, 'I-95 Corridor Coalition 16 Vehicle Probe Project', University of Maryland, College Park. Lahrmann, H, Skoven Pedersen, K & Christensen, LT, 2010, 'Bluetooth detektorer som ny cost-effektiv sensor i vejtrafikken', Selected Proceedings from the Annual Transport Conference at Aalborg University. Malinovskiy, Y., Wu, Y.J., Wang, Y., Lee, U.k. 2010 'Field Experiments on Bluetooth-Based Travel Time Data Collection'. TRB 89th Annual Meeting Compendium of Papers DVD. Wasson, J.S., J.R. Sturdevant, and D.M. Bullock. 'Real-Time Travel Time Estimates Using Media Access Control Address Matching'. ITE Journal, Vol. 78, No. 6, 2008, pp. 20-23. Puckett D. D, Vickich M. J, 2010 'Bluetooth-Based Travel Time/Speed Measuring Systems Development' Texas Transportation Institute, UTCM Project #09-00-17. Tarnoff, P.B., Darcy; Young, Stanley; Wasson, James; Ganig, Nicholas; Sturdevant, James. 2009 'Continuing Evolution of Travel Time Data Information Collection and Processing'. Transportation Research Board 88th Annual Meeting, Washington D.C. CD-ROM.