A Low-Cost, Human-Inspired Perception Approach for Dense Moving Crowd Navigation

Abstract

Robot travel through dense moving crowds is necessary for timely execution of navigation tasks since humans and other robots depend on reliable arrival times. Therefore, robots need socially appropriate methods for finding and entering openings in free-flowing crowds. We describe a method for low-cost awareness of group formation, personal space approximation, and occlusion compensation for use in crowd navigation. Our approach attempts to incorporate social expectations and mimic human techniques when traversing dense, moving crowds. Humans usually cluster people into groups of similar speed and direction, instead of tracking individual motion, and take into account partially obscured people. Using a similar approach avoids the challenges of detecting and tracking a large number of people simultaneously. Another human technique is to estimate and gamble to reduce perceptual and cognitive load. Our approach uses a single RGB-D sensor and (1) clusters all moving objects into groups and detects splits and merges in these groups, (2) applies a 2D polygon projection in obscured regions to reduce inappropriate motion and collisions due to unexpected concealed objects, and (3) defines a dynamic group personal space modeled using asymmetric Gaussians in order to inhibit certain socially inappropriate robot paths. This approach trades off detection of individual people for higher coverage and lower cost, while preserving high speed processing. A real-world quantitative evaluation of this approach showed good performance in comparison to an existing people detection approach. The projected polygon step captures significantly more people in the scene (77% vs. 80%) and supports group clustering in dense, complex scenarios. In addition to quantitative findings, there were multiple interesting examples cases of group characteristics and behaviors where other approaches typically encounter difficulty.