Crowd Segmentation Through Emergent Labeling

Abstract

As an alternative to crowd segmentation using model-based object detection methods which depend on learned appearance models, we propose a paradigm that only makes use of low-level interest points. Here the detection of objects of interest is formulated as a clustering problem. The set of feature points are associated with vertices of a graph. Edges connect vertices based on the plausibility that the two vertices could have been generated from the same object. The task of object detection amounts to identifying a specific set of cliques of this graph. Since the topology of the graph is constrained by a geometric appearance model the maximal cliques can be enumerated directly. Each vertex of the graph can be a member of multiple maximal cliques. We need to find an assignment such that every vertex is only assigned to a single clique. An optimal assignment with respect to a global score function is estimated though a technique akin to soft-assign which can be viewed as a form of relaxation labeling that propagates constraints from regions of low to high ambiguity. No prior knowledge regarding the number of people in the scene is required.