Color-based 3D particle filtering for robust tracking in heterogeneous environments

Most multi-camera 3D tracking and positioning systems rely on several independent 2D tracking modules applied over individual camera streams, fused using both geometrical relationships across cameras and/or observed appearance of objects. However, 2D tracking systems suffer inherent difficulties due to point of view limitations (perceptually similar foreground and background regions causing fragmentation of moving objects, occlusions, etc.) and, therefore, 3D tracking based on partially erroneous 2D tracks are likely to fail when handling multiple-people interaction. In this paper, we propose a Bayesian framework for combining 2D low-level cues from multiple cameras directly into the 3D world through 3D particle filters. This novel method (direct 3D operation) allows the estimation of the probability of a certain volume being occupied by a moving object, using 2D motion detection and color features as state observations of the particle filter framework. For this purpose, an efficient color descriptor has been implemented, which automatically adapts itself to image noise, proving able to deal with changes in illumination and shape variations. The ability of the proposed framework to correctly track multiple 3D objects over time is tested on a real indoor scenario, showing satisfactory results.