Real-time depth-based tracking using a binocular camera
Leijie Zhang, Zhiqiang Cao, Xiangrui Meng, Chao Zhou, Shuo Wang · 2016
Depth map provides rich information and it can be utilized in object tracking to handle some challenging problems in conventional RGB tracking such as occlusions and model drift. In this paper, we present a tracker that provides an effective real-time target tracking method based on a binocular camera. The proposed tracker is an extension of the popular KCF algorithm that leverages a circulant structure of tracking-by-detection with kernels for tracking. On this basis, we design a simple yet effective method to detect occlusions and recover tracking using noisy depth map obtained from a binocular camera. Firstly, one needs to identify exact and reasonable peak number of the target region's depth histogram and apply GMM model to evaluate the depth distribution. The occlusion is then detected based on the depth evaluation and the maximum response from KCF. When the occlusion happens, it is segmented to build the corresponding search region for recovery. The experimental results demonstrate the effectiveness of the proposed method and the superiority to the KCF tracker.