Traffic Accident Anticipation via Driver Attention Auxiliary

Traffic accident anticipation in driving videos aims to provide early warning of accidents and encourage accurate decision-making. Previous research has primarily focused on the spatial temporal correlation at the object level, but it lacks some explainable clues and is susceptible to severe environmental changes. Hence we propose a method that utilizes driver attention as an auxiliary factor for traffic accident anticipation (DA-TAA) to enhance model training in this work. Specifically, driver attention provides valuable insights into key areas closely related to safe driving. DA-TAA consists of a self-attention feature extraction module, a temporal GRU module, and a driver attention-guided accident prediction module. We employ attention mechanisms to explore driver attention cues for accident prediction. We train the model using the DADA-2000 dataset, which includes annotated driver attention per frame and evaluate its performance on both the DADA-2000 and CCD datasets. Our extensive experiments demonstrate that DA-TAA outperforms state-of-the-art methods in traffic accident anticipation.