Static Occlusion Detection and Handling in Transportation Videos

Occlusions present a challenge in surveillance and traffic monitoring applications where person and/or vehicle tracking are required. Video-based object tracking is a process where the location of a given object of interest in a video sequence is determined across a range of frames. A key step in typical tracking operations is forming a feature representation of an object being tracked and solving a correspondence problem to find the location of the best-matching set of those features between video frames. The best-matching feature set is usually found via optimization algorithms across regions in subsequent frames near and around the location of the object in a current frame. The features used to solve the correspondence problem are usually appearance-based, and may include color, texture and shape descriptors. Consequently, the track can be lost when a view of the tracked object is occluded by objects in the scene because the appearance of the occluded object may not sufficiently resemble the appearance of the unoccluded object. In this paper, we present a method for determining the location of static occlusions in a scene at the pixel level, and utilize the knowledge of the location of the occlusions to boost the performance of well-known video-based object tracking algorithms. We demonstrate via experimental testing that the proposed method is effective in improving the performance of tracking algorithms, particularly when the motion in the scene is highly regularized, as is the case in cameras performing transportation monitoring tasks.