A Multi-Module Perception-Based Robot Dynamic Target Following Method—Collaboration of Improved KCF Tracking, Distance Mapping Function, and Obstacle Avoidance

To address visual tracking drift caused by scale variation and occlusion, large errors of monocular ranging, and the insufficient obstacle-avoidance response in robot dynamic target following, this paper proposes a multimodal perception–based following framework. First, a KCF tracker is enhanced with a closed-loop scale-adaptive mechanism and a four-level occlusion handling scheme, so that robust target localization can still be maintained under severe appearance changes. Second, a camera-calibration–driven distance mapping function is constructed on top of the perspective projection model by introducing geometric distortion compensation, pose compensation, and a learnable bias term, which brings the monocular distance error down to the centimeter level within the 1–4 m working range. Third, LiDAR point cloud is used to build an obstacle threat model, and a semantic TEB local planner is employed to achieve coordinated target following and obstacle avoidance on a CPU-only mobile robot platform. Experiments on public long-term tracking benchmarks (LaSOT, TrackingNet, GOT-10k) and on an indoor TurtleBot3 platform show that, compared with the original KCF, the proposed method improves the LaSOT AUC from ~ 20% to 39.1% while keeping real-time performance, and, in real robot scenarios with occlusion, scale change, and dynamic obstacles, achieves a following success rate of 95.2% and a calibrated distance MAE of about 0.027 m (≈2.7%), demonstrating strong practicality and reproducibility.