Deep Spatio-Temporal Memory and Interaction Network for Traffic Incident Detection

In traffic scenes, identifying traffic incidents captured by onboard cameras in driving videos is an essential task. Traffic incidents in driving videos usually have short durations, and the background dynamics are complex, making traffic incident detection a challenging task. To better detect traffic incidents in driving videos, this paper proposes a traffic incident detection method based on a deep spatio-temporal memory and interaction network. First, in this study, a spatio-temporal information perception network is designed that can effectively capture the core semantics from small spatial areas to large spatial areas, as well as short-term features in the temporal domain, to enhance the video frame feature extraction ability and provide a more accurate classification for traffic incidents. Second, this paper proposes a temporal feature learning network so that the spatio-temporal information perception network can fully capture long-term temporal information when processing video frames. This method enhances the learning of time features to obtain more complete contextual temporal information features and improve the accuracy of determining the abnormal occurrence and termination times. To evaluate our method, we conducted extensive experiments on the DoTA dataset containing 4677 videos with temporal, spatial, and categorical annotations. Our experimental results show that our method achieves an AUC of 80.9% for traffic scenes. These results are highly competitive when compared with the existing state-of-the-art methods, demonstrating good traffic incident detection classification accuracy and efficiency.