Detecting and Tracking the Moving Vehicles Based on Deep Learning

Aiming to address the problem of counting multi-target moving vehicles in the various complex traffic environments, this paper proposes a detecting and tracking method based on YOLO (You Only Look Once) and Deep Sort, and evaluates its performance with public dataset (TUA-DETRAC) and two self-collection datasets. The YOLOv4 algorithm is firstly used to detect each moving vehicles, and then Deep Sort algorithm is adopted to track multi-target vehicles. The experimental results show that moving vehicles can be effectively detected and tracked in real time under different traffic environments including daytime, nighttime, rainy and crowded scenes. The experimental results show that the proposed method can reach 93% average detection accuracy with 20fps of tracking speed, and is capable of dealing with different traffic and climate conditions.