Importance sampling as one solution to the data association problem in multi-target tracking

Abstract

Tracking multiple targets with similar appearance is a common task in many computer vision applications as surveillance or sports analysis. We propose a Rao-Blackwellized Resampling Particle Filter (RBRPF) as a real-time multi-target tracking method that solves the data association problem by a Monte Carlo approach. Each particle containing the whole target configuration is predicted by using a process model and resampled by sampling associations and fusing of the predicted state with the assigned measurement(s) instead of the common dispersion. As each target state is modeled as a Gaussian, Rao-Blackwellization can be used to solve some of these steps analytically. The sampling of associations splits the multi-target tracking problem in multiple single target tracking problems, which can be handled by Kalman filters in an optimal way. The method is independent of the order of measurements which is mostly predetermined by the measuring process in contrast to other state-of-the-art approaches. Smart resampling and memoization is introduced to equip the tracking method with real-time capabilities in the first place exploiting the discreteness of the associations. The probabilistic framework allows for consideration of appearance models and the fusion of different sensors. A way to constrain the multiplicity of measurements associated with a single target is proposed and - along with the ability to cope with a high number of targets in clutter - evaluated in a simulation experiment. We demonstrate the applicability of the proposed method to real world applications by tracking soccer players captured by multiple cameras through occlusions in real-time. © 2010 Springer-Verlag Berlin Heidelberg.