A multi-object tracking algorithm based on YOLOv5-concise network
Baoping Cheng, Yan Qun Huang, Xiaoyan Xie, Du Jiawei · 2022
As the basis of pedestrian and traffic statistics, multi-object tracking (MOT) is widely used in smart cities and smart shops. Due to the high cost of cloud analysis, MOT algorithm is mainly deployed on the end-side device. However, limited by computing resources and capital cost, it is difficult to deploy large models in end-side device. As tracking based on detection has become the most effective MOT method, a lightweight detection network model YOLOv5- Concise with only 0.302M parameters is proposed for end-side deployment in this paper. Then, the application of knowledge distillation technology based on outputs and feature relation in object detection is studied. By knowledge distillation, the YOLOv5-Concise model's detection precision mAP0.5 is increased by 7.26%, mAP0.5:0.95 is increased by 16%, and the detection speed remained unchanged. Finally, on the camera equipped with T40 chip, the detection speed of the model is measured to 28 FPS, and the accuracy rate of 98.1% is obtained in the pedestrian statistics test.