Deep learning classification applied to traffic accidents prediction

Abstract

In this paper, YOLOv4 neural networks are trained with the goal of detecting and classifying objects from a street as seen from a drone. These have been trained on the VisDrone dataset, which is firstly validated through a custom-made graphic user interface. Then, several tests regarding performance, dataset composition and contrast have been carried out on the trained models. Results are compared to those from other previously existing models in order to evaluate their performance and analyse their shortcomings. The trained models are then used to detect and classify objects in a city scenario in real-time. Finally, an algorithm is proposed to track the objects, infer their future trajectories and predict potential collisions from the expected trajectories.