AAT-DA: Accident Anticipation Transformer with Driver Attention

Traffic accident anticipation is an important issue for reducing the number of road fatalities and for realizing safe autonomous driving. Conventional models for accident anticipation primarily rely on Recurrent Neural Networks (RNNs) and have achieved notable success. In contrast, transformers have recently achieved significant success in various fields, including video processing. However, their performance in accident anticipation remains below that of RNN-based models. In this study, we propose the Accident Anticipation Transformer with Driver Attention (AAT-DA), a novel model that leverages transformers for both temporal and spatial feature extraction. The model also leverages driver attention to focus on objects likely to be involved in an accident. Additionally, the model can specify an object that moves dangerously through the attention matrix. The model recorded the state-of-the-art anticipation performance on two representative datasets for accident anticipation.