Using multi-view video to overcome the occlusion problem in object tracking

Abstract

Object tracking is still a challenging problem despite it being a crucial component of many computer vision applications. Even with the tremendous advancements in object tracking algorithms, object trackers still often fail in the presence of illumination changes, background clutter and occlusions. Occlusions pose a serious challenge as even if the object of interest is only occluded for a short time, it can result in the tracker losing track of the object entirely. The additional information provided by 3D video content, namely the additional views of the scene and the additional depth information, provide an opportunity for solving the occlusion problem. This work aims to solve the occlusion problem by utilising the additional information provided by 3D video content. This work experiments with various techniques such as Kalman filtering and machine learning when employed on 3D video. A new occlusion aware tracking system is developed, which employs a 3D Kalman filter to model the object’s motion. The tracking system demonstrated an improvement in tracking accuracy and robustness on some sequences. In the sequences in which the tracking solution fails, it fails in a similar way to the state art solutions.