A Robust and Efficient Approach for Human Tracking in Multi-Camera Systems

In this paper, a robust and efficient approach for multi-camera human tracking is presented. The approach is integrated in an experimental surveillance system, based on a camera network with a task-oriented architecture. At sensor level, image processing algorithms are applied for object detection and feature extraction. Additionally, for each object that is to be tracked, an agent-based multi-sensor process is created, which autonomously performs multi-sensor data association and fusion. One of the major challenges in such systems is to robustly determine correspondences between observations from different sensors with different environmental conditions. Therefore, in this paper, efficient and robust spacial and appearance features for object description and recognition are proposed. For spacial description an approximated object position in world coordinates is estimated and evaluated by an inconsistency detector before associated to a Kalman Filter. For appearance similarity calculation, an appearance model is proposed and a similarity metric based on the Earth Mover's Distance (EMD) is presented. Finally, the data fusion algorithm based on these features for tracking objects in overlapping and non-overlapping camera networks is presented.