Human Target Detection and Tracking Algorithm Based on Improved YOLOv7

Abstract

Recently, human-oriented target tracking has broad application prospects in various practical fields, such as medicine and security. YOLOv7 is a powerful detection algorithm in the current main-stream, and it still has room for development in terms of detection accuracy. To handle the issue of mobile robots struggling to maintain recognition precision and stability in complicated scenes, we propose an improved target detection network for YOLOv7. Firstly, based on YOLOv7-tiny network structure, the CBAM attention mechanism is introduced into the bottom of the backbone network, and the SPPCSP module is adopted to reduce the computation of the feature map, which could enhance the network's representational capability. Then, the KCF algorithm is incorporated into the model, which could track the human target after determining its position with respect to the robot with the RGB-D camera. In order to ensure the flexibility of the robot, the Jetson Nano is utilized as the processor. Finally, we conduct human target tracking experiments for the mobile robot in a realistic scenario. The experimental results show that the mAP reaches 91.17%, which is 1.88% higher than only using YOLOv7 after introducing the CBAM module.