Real-Time Multi-scale Tracking via Online RGB-D Multiple Instance Learning

Abstract

It is still a challenging problem to develop robust target tracking algorithm under various environments. Most of current target tracking algorithms are able to track objects well in controlled environments, but they usually fail in significant variation of the target's scale, pose and plane rotation. One reason for such failure is that these object tracking algorithms employ fixed-size tracking box, and the other is that traditional 2D feature-based tracking algorithms are lack of 3D information. In this paper, we address the two problems by combining the fused 3D features and the bootstrap filter. So a multi-scale RGB-D tracker is proposed. The multi-scale RGB-D tracker has several attractive merits: (1) It exploits multi-instance learning strategy for fusing effective 3D information from the RGB image and the corresponding depth image. (2) It uses the bootstrap filter to solve the problem of target losing, when target changes significantly in scale. (3) Our tracking algorithm also reduces the error accumulation and can obtain a good performance when the candidate target is not so well. Extensive experiments demonstrate the effectiveness of the proposed tracking algorithm in indoor and outdoor environments where the targets undergo large changes in pose, scale, and plane rotation.