Occlusion handling method for object tracking using RGB-D data

We propose a novel method for occlusion handling in object tracking using RGB-D sequences. The proposed method is designed as a stand-alone system which can be integrated into any 2D tracker, and consists of three stages: 1) Occlusion detection by fitting a Gaussian Mixture Model (GMM) to the depth distribution of the targets bounding box. 2) Reliable tracking during partial and full occlusions by incorporating RGB and depth segmentation of the scene. 3) Target recovery using depth information and motion estimation. We carry out a quantitative evaluation which shows a significant improvement in performance of state of the art RGB trackers, and also in comparison to another RGB-D tracker.