Particle Filtering with a Soft Detection Based Near-Optimal Importance Function for Visual Tracking
Abstract
Particle filters are currently widely used for visual tracking. In order to improve their performance, we propose to enrich the observation model with soft detection information and to derive a near-optimal proposal to efficiently propagate particles in the state space. This information reflecting probabilities about the object location is more reliable than the usual binary output which can yield false or missed detections. Moreover, our proposal not only incorporates the observations as in previous works, but relies on a close approximation of the optimal importance function. The resulting PF achieves high tracking accuracy and has the advantage of coping with unpredictable and abrupt movements.