A Complementary Visual Tracking Framework Based On Mid-level Visual Cuses And High-level Structure Information

The research on Object Tracking is always an important field in computer vision. These technologies are highly developed during the past decade. For the reason that the technique of object tracking models can be widely expended in various fields for both civilian use and military use, for instance, crowd detection, subway security monitoring, indoor security surveillance, human stream estimation, pedestrian monitoring and estimation, and so forth, it has drawn wide attentions and has a extensive prospect for future applications.At first, this paper selects and summarizes up some basic theories, classic algorithms and the shortcomings and problems of existed methods on the relative fields of visual object tracking. Then, we generalize the challenging factors and recent focuses of the object tracking field. Next, we come up with a novel concept on visual tracking, which is totally different from former algorithms in the past. Afterwards, we construct up a more robust and reasonable visual tracking framework based on different levels of visual cues. Extensive experiment has shown that, compared with the classic methods in the past, the algorithm proposed by this paper is much more competent in solving the most challenging factors which the visual tracking field is faced with nowadays.While numerous algorithms have been proposed for object tracking with demonstrated success, it remains a challenging problem for a tracker to handle large change in scale, motion, shape deformation with occlusion. One of the main reasons is the lack of effective image representation to account for appearance variation. Most trackers use high-level appearance structure or low-level cues for representing and matching target objects. In this paper, we firstly propose a novel appearance model, which is called object tracker, for representing a target in tracking task from the perspective of mid-level vision with structural information captured in superpixels. We present a discriminative appearance model based on superpixels, thereby facilitating a tracker to distinguish the target and the background with mid-level cues. For this object tracker of novel appearance model, the tracking task is then formulated by computing a target-background confidence map, and obtaining the best candidate by maximum a posterior estimate.Furthermore, we design a region tracker based on high-level structure information (incremental PCA), and adopt multi-state particle filter to integrate these two complementary trackers, the object tracker and region tracker, into a robust tracking framework. While region tracker is more robust to scaling and in-plane rotation, object tracker is more competent in dealing with out-of-plane rotation and deformation. Experiment shows that these two tracker supervise each other against different challenges, and our Complementary Visual Tracking (CVT) framework can resist scaling, deformation, in-plane rotation and out-of-plane rotation simultaneously.