Attention Based Framework for Traffic and Collision

This paper presents a solution aimed at mitigating the risk of traffic collisions and automating immediate emergency response when such incidents occur. Our approach involves cascading an object detection model, supplemented with a tracking algorithm and a prediction model based on attention. To select the appropriate detector a comparison between recent proposals is implemented. A secondary network is then employed to forecast the future positions of traffic participants. With these predictions, we calculate the potential for contact among the objects, providing a predictive measure of collision risk. The proposed solution is evaluated on a relevant database and the best performers are identified.