Tracking of objects in 3D video content

Abstract

Object tracking is a critical component in many computer vision applications. Despite years of research it is still considered a difficult problem. Stereo and multi-view cameras present new capturing techniques which provide extra information about a scene, such as extra views and the depth of objects in the scene. This information can be exploited to increase tracking accuracy and allow for more robust object trackers which do not fail in situations of clutter, similarities in appearance between the object and the background, and partial occlusions. In this work a 3D object tracking system which builds upon the widely used mean-shift tracking algorithm was designed and implemented. The emphasis is on ease of use, with the parameters required being determined automatically without user intervention, and on efficiency in terms of speed. A significant increase in tracker robustness was observed. The 3D tracker successfully tracked objects, which are similar in appearance to the background and clutter, throughout the entire video sequence, whereas the 2D counterpart failed to track the objects completely. The performance of the 3D tracker, in terms of speed, is on par with the 2D mean-shift tracker, and is suitable for real-time use.