A Simulation-Based Framework for Urban Traffic Accident Detection

Urban traffic accident detection in surveillance videos is an essential task in intelligent traffic monitoring. The current deep-learning based approaches suffer from the lack of large-scale traffic accident video datasets due to the privacy concerns. In this paper, we propose a framework to synthesize traffic videos containing both normal traffic and accident events by simulating the real urban traffic scenarios. By introducing various weather conditions, viewpoints of surveillance cameras, and traffic statuses, we can simulate the complex real-world traffic scenarios. The vehicles are automatically controlled to behave normally or trigger accidents randomly. Based on our framework, we automatically generate a synthetic dataset containing 1,100 distinct traffic videos with video-level annotations. For accident detection, we observe that the existing reconstruction-based models produce many false alarms by predicting the pseudo accidents (non-accident anomalous traffic events) as real accidents. To address this ambiguity problem, we propose an improved approach by employing multiple instance learning to refine the detection results. Our experiments demonstrate the great potentials of our proposed framework in boosting the performance of accident detection on real-world datasets. We make our code and dataset publicly available at https://github.com/hankluo2/UrbanTrafficAccidentDetection.