Light-Weight Multi-Target Detection and Tracking Algorithm Based on M3-YOLOv5

The existing multi-target detection and tracking algorithms network models are too large to be deployed on small mobile devices. In order to solve the above problems, a light-weight multi-target detection and tracking algorithm based on M3-YOLOv5 is proposed. Firstly, the YOLOv5s backbone network is replaced by improved MobileNetV3 with sandglass block to reduce the algorithm size. Secondly, the CA attention mechanism is combined to the neck network to improve the small and medium-sized targets detection capability. Finally, DeepSORT is combined to track the obtained detection border information. The application experiment on the NVIDIA Jetson TX2 development board shows that YOLOv5s inference speed is accelerated by 24.6%. The improved multi-target tracking algorithm inference speed is accelerated by 2FPS in comparison with DeepSORT. The algorithm results in less loss of tracking accuracy and enables multiple targets to be tracked quickly, which proves to be suitable for the deployment on small mobile devices.