Deformation Gated Recurrent Network for Lane-Level Abnormal Driving Behavior Recognition

Guojiang Shen,Juntao Wang,Xiangjie Kong,Zhanhao Ji,Bing Zhu,Tie Qiu
DOI: https://doi.org/10.1145/3635141
2024-01-01
Abstract:As a significant part of traffic accident prevention, abnormal driving behavior recognition has been receiving extensive attention. However, the granularity of existing abnormal driving behavior recognition is mostly at road-level, and these methods’ high complexity leads to high overhead on training and recognition. In this article, we propose a deformation gated recurrent network for lane-level abnormal driving behavior recognition. Firstly, we use conditional random field model to calculate the lane change necessity of the vehicle, which helps us to distinguish whether the lane-changing behavior is reasonable. Secondly, we propose deformation gated recurrent network (DF-GRN) and trajectory entropy to capture the implicit relationship between trajectories and shorten recognition time. Finally, we get classified results including aggressive, distracted and normal driving behavior from the network. Distracted and aggressive behavior will be marked as anomaly. The effectiveness and real-time nature of the network are verified by experiments on Hangzhou and Chengdu location datasets.
What problem does this paper attempt to address?