A Compute-Efficient Human-Following Robotic Mobility System Integrating Multimodal Recognition and Hierarchical Local Planning

This paper presents a compute-efficient human-following robotic mobility system that integrates multimodal target recognition, LiDAR-based obstacle perception, hierarchical local planning, and low-level motion control into a unified perception–planning–control architecture. The system enables real-time target tracking and obstacle avoidance on resource-constrained embedded hardware. RGB-D vision and Tether Follow Sensors (TFS) are used for target recognition, while LiDAR provides local obstacle information. To reduce the computational burden of local planning, a Hierarchical Dynamic Window Approach (HDWA) is proposed. Unlike conventional DWA, which uniformly evaluates sampled motion candidates, HDWA applies Movement, Direction, and Detail dynamic windows according to obstacle conditions and target direction. This hierarchical structure reduces redundant trajectory evaluations by activating additional computation only when avoidance is required. Experimental validation on a low-cost embedded platform demonstrates that HDWA reduces the candidate-evaluation workload by 30.8–92.3% while maintaining stable target tracking and safe obstacle avoidance. In the left- and right-side avoidance scenarios, HDWA also reduced the path length by 6.5–7.8%, increased the average driving speed by 22.0–30.8%, and reduced the elapsed motion time by 24.4–28.5%. These results demonstrate the feasibility of the proposed compute-efficient system for human-following robots operating under embedded hardware constraints.