Mobile Robot Tracking Method Based on Improved YOLOv8 Pedestrian Detection Algorithm

Abstract

Human following of mobile robot is a classical application field in the category of human-computer interaction. To address the problem of unstable following of the target in complex scenes, a method of moving robot target following (YOB-CF) based on YOLOv8CSM-tiny and BKPs-CF (Correlation Filtering of Body Key Points) algorithm is proposed. Firstly, based on the YOLOv8 Network, the backbone network is replaced by Multi-Head Attention under this framework, which focuses on improving the detection accuracy of human target recognition in complex scenes, while lightening the model, the model reasoning speed and parameters are reduced to meet the deployment requirements of mobile robot platform. Then the BKPs-CF algorithm based on kernel correlation filter (KCF) is used to fuse the features of body key points, and the key points are updated by optical flow method under the condition of target occlusion, to ensure that the robot can still track the target stably under the occluded condition. Finally, the quantitative experiments are carried out on the self-built test set and the public dataset, and the target-following test and performance evaluation of the mobile robot are completed in the real scene, experimental results show that the proposed method is robust and real-time.