Traffic Anomaly Detection via Perspective Map based on Spatial-temporal Information Matrix

Shuai Bai, Zhiqun He, Yu Lei, Wei Wu, Chengkai Zhu, Ming Sun, Junjie Yan; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 117-124

Anomaly detection on the road traffic has vast application prospects in urban traffic management and road safety. Due to the impact of many factors such as weather, viewpoints and road conditions in the real-world traffic scene, anomaly detection still faces many challenges. There are many causes for vehicle anomalies, such as crashes, vehicle on fires and vehicle faults, which exhibits different unknown behaviors. In this paper, we propose an anomaly detection system that includes three modules: background modeling module, perspective detection module, and spatial-temporal matrix discriminating module. The background modeling analyses the traffic flow to obtain the road segmentation results, and the vehicle flow superposition is used to obtain the continuous stationary region. The perspective detection module gets the perspective map by the first detection result, through which the image is cropped to uniform scale for different vehicles and re-detection. Finally, we get all anomalies by constructing spatial-temporal information matrix with the detection results. Furthermore, all anomalies are merged through the non maximum suppression (NMS) and the re-identification model, including spatial and temporal dimensions. The experimental results show that our system is effective in the Track3 test-set of NVIDIA AI CITY 2019 CHALLENGE, which finally ranked first in the competition, with a 97.06% F1-score and 5.3058 root mean square error (RMSE).