Multi-View Classifier Swarms for Pedestrian Detection and Tracking

We describe a novel method for recognition and localization of objects in 3D space using multiple views. We pose the task of classifying and locating objects in 3D space as an optimization problem that combines 2D classifier scores from two separate views of the object. Our methods combine feature-based 2D object classification with efficient search mechanisms based on Particle Swarm Optimization (PSO) by implementing each particle as a local window classi- fier guided by PSO dynamics. We first formulate the localization problem in a multi-view framework where the domain of swarm classifier particles is extended from 2D image space to 3D space. In this scheme, swarm particles encapsulate the geometric structure that binds multiple views. Each particle is a self-contained classifier that "flies" through the solution space seeking the most "objectlike" region in space, thus effectively expanding the scope of an image based 2D classifier to 3D space. In a second formulation, we demonstrate how 2D swarm based search results obtained from individual 2D views can be used to perform 3D localization tasks. We outline a method to extend the framework from single object to multiple object localization and tracking scenarios.