Robust 3D multi-camera tracking from 2D mono-camera tracks by Bayesian association

Visual tracking of people is essential automatic scene understanding and surveillance of areas of interest. Monocular 2D tracking has been largely studied, but it usually provides inadequate information for event interpretation, and also proves insufficiently robust, due to view-point limitations (occlusions, etc.). In this paper, we present a light but automatic and robust 3D tracking method using multiple calibrated cameras. It is based on off-the-shelf 2D tracking systems running independently in each camera of the system, combined using Bayesian association of the monocular tracks. The proposed system shows excellent results even in challenging situations, proving itself able to automatically boost and recover from possible errors.