Eye-SCAN: Eye-Movement-Attention-based Spatial Channel Adaptive Network for traffic accident prediction

Abstract

In the task of using visual cues extracted from DashCam video data to predict future accidents, understanding the dynamic spatio-temporal interactions in driving scenarios poses a major challenge. Given that the gaze attention information of experienced drivers during the driving process involves complex spatio-temporal interactions, this information can provide valuable guidance for training accident prediction models. Therefore, we propose an Eye-Movement-Attention-based Spatial Channel Adaptive Network (Eye-SCAN) for traffic accident prediction, which can efficiently learn multi-scale spatial channel information from driver gaze data. To integrate potential guidance information from driver eye movement information (EyeInfo) into Eye-SCAN, we propose two sub-modules in our model: the Spatial Adaptive Module (SAM), which helps Eye-SCAN adaptively learn low-dimensional spatial features of EyeInfo; and the Channel Adaptive Module (CAM), which aids Eye-SCAN to adaptively learning high-dimensional channel features of EyeInfo. Additionally, we introduce a novel recursive transmission strategy for temporal information to mitigate the impact of varying past results on the model’s current inferences. Experimental results demonstrate that our model outperforms state-of-the-art methods on two benchmark datasets, highlighting the contributions of each component and offering an effective solution for enhancing the safety of intelligent vehicles.