Real-Time Pedestrian Tracking in Natural Scenes

In computer vision real–time tracking of moving objects in natural scenes has become more and more important. In this paper we describe a complete system for data driven tracking of moving objects. We apply the system to tracking pedestrians in natural scenes. No specialized hardware is used. To achieve the necessary efficiency several principles of active vision, namely selection in space, time, and resolution are implemented. For object tracking, a contour based approach is used which allows contour extraction and tracking within the image frame rate on general purpose architectures. A pan/tilt camera is steered by a camera control module to pursue the moving object. A dedicated attention module is responsible for the robustness of the complete system. The experiments over several hours prove the robustness and accuracy of the whole system. Tracking of pedestrians in a natural scene has been successful in 79% of the time.