Detecting Anomalous Bus-Driving Behaviors from Trajectories

Abstract

In urban transit systems, discovering anomalous bus-driving behaviors in time is an important technique for monitoring the safety risk of public transportation and improving the satisfaction of passengers. This paper proposes a two-phase approach named Cygnus to detect anomalous driving behaviors from bus trajectories, which utilizes collected sensor data of smart phones as well as subjective assessments from bus passengers by crowd sensing. By optimizing support vector machines, Cygnus discovers the anomalous bus trajectory candidates in the first phase, and distinguishes real anomalies from the candidates, as well as identifies the types of driving anomalies in the second phase. To improve the anomaly detection performance and robustness, Cygnus introduces virtual labels of trajectories and proposes a correntropy-based policy to improve the robustness to noise, combines the unsupervised anomaly detection and supervised classification, and further refines the classification procedure, thus forming an integrated and practical solution. Extensive experiments are conducted on real-world bus trajectories. The experimental results demonstrate that Cygnus detects anomalous bus-driving behaviors in an effective, robust, and timely manner.