Improved Dynamic Spatial-Temporal Attention Network for Early Anticipation of Traffic Accidents

The proliferation of dashcams has significantly increased the volume of recorded data on road traffic, offering a unique opportunity to develop sophisticated algorithms that can analyze this data and predict potential accidents. In this paper, we introduces a new approach for predicting car accidents in videos by leveraging the Dynamic Spatial-temporal Attention Network (DSTA). We incorporate a frame-level loss and a bag-level loss into our method to aid in the model's learning process. Moreover, given that car crashes often involve continuous processes, we introduce soft labels to smoothen the label transitions in each video frame, thereby helping the model to more accurately identify the accident frame number and minimizing the impact of labeling noise. Additionally, we employ the output of the temporal self-attention aggregation (TSAA) module to enhance the prediction's robustness, which extracts information from all frames of the current video and to avoid interference from individual difficult frames. Our experimental results indicate that our Improved-DSTA (IDSTA) method outperforms the original DSTA method and performs exceptionally well in the AVA dataset. Overall, our proposed approach demonstrates significant potential in predicting car accidents from dashcam videos.