Learning Sparse Prototypes via Ensemble Coding Mechanisms for Crowd Perception

This paper presents a novel approach to learning a dictionary of crowd prototypes for dynamic visual scenes. Recent work in cognitive psychology suggests that crowd perception may be based on pre-attentive ensemble coding mechanisms [24] in the spirit of feedforward hierarchical models of visual processing [4]. We extend a biological model of motion processing [10] with a new dictionary learning method tailored for crowd perception. Our approach learns crowd prototypes through explicit ensemble coding mechanisms via structural and local coherence constraints. We evaluate the proposed method on multiple crowd perception problems from collective or abnormal crowd detection to tracking individuals in crowded scenes. Experimental results on crowd datasets demonstrate competitive results on par or better than the state of the art.