Robust visual tracking based on adaptive depth-color-cue integration using range sensor

In visual tracking field, multi-cue integration has been researched extensively, but only color-based method still suffers from illumination changes, color-similar background or complete occlusion. To overcome these shortages, this paper presents an adaptive depth-color-cue integration framework for Mean-shift tracking. The state-of-art 2D rectangles evolves to 3D cubes for representing target region, and depth and color cues are combined together for representing target appearance. Moreover, a novel depth-data-based motion detection method is introduced to get more reliable motion cues during tracking. Furthermore, a reliability evaluation function is proposed to tune cues' weights based on the assumption that most reliable cues are those which are most discriminative between target region and background regions. Finally, cues' probability distribution maps are integrated for Mean-shift tracking. Extensive experiments under various conditions demonstrate the reliability and robustness of this depth-color-integrated tracking framework.