Using Machine Learning to Estimate Pedestrian and Bicyclist Count of Intersection by Bluetooth Low Energy

Abstract

Quantifying pedestrian and bicycle traffic is important for planning, investment, and safety improvements. Traffic agencies have implemented various pedestrian/bicyclist detection systems, but the accuracy is unsatisfactory for intersections. Some studies have explored the use of media access control (MAC) address-scanning sensors such as Bluetooth and Wi-Fi scanners. However, they may suffer from low detection rates. To overcome these shortcomings, this study proposed a system based upon Bluetooth low energy (BLE) scanners. First, the feasibility was assessed by identifying the detection rate and range of BLE scanners. Evaluation experiments uncovered that the detection rate is much higher than the Bluetooth ordinary, and it is sufficiently high for traffic count studies. Moreover, the detection range could cover the whole intersection while reducing the overestimating caused by the large detection range in comparison with other MAC address–scanning sensors. A two-step framework is then proposed for identifying the pedestrians and bicyclists from stationary objects and motorized travelers using one of the popular machine-learning algorithms, one-class support vector machine. The proposed system is validated by the benchmark count data from video footage. The results show that the system can reasonably estimate the counts of pedestrians and bicyclists in a mixed-traffic environment. The average absolute percentage error is 6.35%. This study has concluded that compared to traditional Bluetooth and Wi-Fi, BLE is more suitable for estimating the counts of pedestrians and bicyclists.