概要
This paper presents a novel collision avoidance approach for mobile robots that considers pedestrians’ visual attention states when generating cost maps. The system integrates visual attention estimation using head pose analysis through MediaPipe framework with ncertainty-aware trajectory prediction based on Monte Carlo Dropout. By detecting distracted states, particularly smartphone usage while walking, the system dynamically adjusts prediction uncertainty bounds and corresponding cost map generation. The effectiveness of the proposed method was evaluated through experiments using a modified electric wheelchair platform in an indoor environment. Results demonstrated that during smartphone use scenarios, our method maintained an average minimum distance of 1.57m compared to 1.31m for the conventional obstacle-based approach, while showing comparable completion times. This integration of attention states with trajectory uncertainty enables more nuanced path planning, leading to safer navigation around distracted pedestrians without significantly compromising efficiency.
