PathCluster: Pedestrian Group-Adaptive Social Navigation in Dense Crowds

Abstract

Mobile robot navigation in crowded spaces is crucial for deployment but remains challenging in extremely dense environments. While recent works have utilized predictive measures to anticipate individual trajectories, bettering overall navigation, these methods often fail to scale in high-density crowds due to computational intensity. To mitigate this issue we propose PathCluster, a novel approach that generates groups based on individuals with similar trajectories to enable efficient navigation in dense crowds. Our method introduces a group generator algorithm that identifies and treats clusters as cohesive units, significantly decreasing the complexity of trajectory prediction while maintaining its benefits. Simulation results demonstrate that our method, PathCluster, achieves a 45% higher success rate, and a 25% lower collision rate, and can tackle more challenging navigational tasks within a 48hr-time limit compared to previous social navigation models in extremely crowded environments.