Real-Time RGB-D Visual Tracking With Scale Estimation and Occlusion Handling

Abstract

Robust scale estimation and occlusion handling are two challenging problems in visual target tracking. Most existing methods do not perform well with complex image sequences, especially those with scale changes and partial occlusion. In this paper, a real-time RGB-D object tracker is proposed to deal with occlusion and scale changes in various scenes. Our approach is more accurate than most existing algorithms, while the computational complexity does not increase greatly. We build our approach based on the kernelized correlation filter (KCF), which utilizes the property of a circulant matrix and kernel to achieve fast target tracking. Our approach makes use of the spatial continuity of the depth information to achieve accurate scale estimation, and it performs better than an exhaustive search. In addition, we propose a method for model updating that combines the different parts of the target to deal with occlusion. We evaluate our method using the Princeton Data, which is a public RGB-D dataset for object tracking. Experiments show that the proposed approach significantly improves the performance compared with the baseline. Finally, we provide both quantitative and qualitative comparisons between our tracker and current state-of-the-art trackers. The proposed approach is shown to be superior to most trackers, while operating at a high frame rate.