TempoLearn Network: Leveraging Spatio-Temporal Learning for Traffic Accident Detection

Recognizing traffic accident events in driving videos is a challenging task and has emerged as a crucial area of interest in autonomous driving applications in recent years. To ensure safe driving alongside human drivers and anticipation of their behaviors, methods to efficiently and accurately detect traffic accidents from a first-person viewpoint must be developed. This paper proposes a novel model, named the TempoLearn network, which leverages spatio-temporal learning to detect traffic accidents. The proposed approach incorporates temporal convolutions, given their effectiveness in identifying abnormalities, and a dilation factor for achieving large receptive fields. The TempoLearn network has two key components: accident localization, for predicting when the accident occurs in a video, and accident classification based on the localization results. To evaluate the performance of the proposed network, we conduct experiments using a traffic accident dashcam video benchmark dataset, i.e., the detection of traffic anomaly (DoTA) dataset, which is currently the largest and most complex traffic accident dataset. The proposed network achieves excellent performance on the DoTA dataset, and the accident localization score, measured in terms of AUC, is 16.5% higher than that of the existing state-of-the-art model. Moreover, we demonstrate the effectiveness of the TempoLearn network through experiments conducted on another benchmark dataset, i.e., the car crash dataset (CCD).