基于局部关系特征和注意力机制的交通事故预测方法

The accident prediction methods with cameras are mainly used to establish the global relationships among traffic objects, lacking the consideration of local relationships among them. Therefore, a predictive model was proposed based on local relational features and attention mechanisms to carry out real-time traffic accident risk prediction with vehicle-mounted cameras. Firstly, a local relational multi-graph network was incorporated to capture the local interactions of vehicles, solving the insufficient application issue of local interaction information about the traffic objects. Subsequently, a dynamic spatial attention mechanism was adopted to identify the risk vehicles at traffic accident. Finally, the Gated Recurrent Unit and dynamic temporal attention mechanism were integrated to effectively utilize the temporal information between the current and historical frames in dynamic scenes. Experimental results on the accident dataset show that the proposed model can predict accident in 1.55 seconds advance with a prediction accuracy of 73.78% and 1.65 ms single-frame prediction time, realizing excellent real-time effectiveness, and providing an effective solution for the traffic accident prediction.