Vision-based accident anticipation with deep neural networks

Abstract

Traffic accidents remain a leading cause of fatalities and injuries worldwide, making the ability to anticipate hazardous situations in advance critical for improving road safety. Accident anticipation systems enable timely preventive actions, such as driver alerts and collision avoidance maneuvers, and play a key role in advanced driver assistance systems (ADAS) and autonomous driving. While many existing approaches rely on multiple sensors such as LiDAR, radar, and GPS, using dash-cam videos offers a cost-effective and easily deployable alternative, albeit with significant challenges due to the complexity of real-world driving environments and real-time constraints. In this thesis, we investigate efficient and accurate accident anticipation from dashcam videos through two complementary approaches. First, we propose an enhanced spatio-temporal graph-based framework (STAGNet), which improves accident prediction by incorporating richer spatio-temporal features and aggregating them using recurrent neural networks. Second, we introduce VAGNet, a novel architecture that leverages global scene representations instead of computationally expensive object-level features, combining transformer and graph modules with the vision foundation model VideoMAE-V2 for efficient feature extraction. Extensive experiments conducted on multiple benchmark datasets, including DAD, DoTA, DADA, and Nexar, demonstrate that the proposed methods achieve superior performance in terms of average precision and mean time-to-accident compared to existing approaches. Notably, the global feature-based framework significantly improves computational efficiency while maintaining high predictive accuracy. Together, these contributions advance the development of scalable, real-time accident anticipation systems using only dash-cam video, providing a practical pathway toward safer intelligent transportation systems.