Robust Tracking With A Structured Local Model

Visual tracking is an important part in the field of computer vision, with the aim of locating the target and capturing the change of target’s motion and appearance in each frame. In this paper, we summarize classic algorithms and theories in visual tracking field. We also comprehensively analyze the history and development trend in this area, the various challenges during handling visual tracking task, and the trend of development in the future, In this paper, we introduce two tracking methods based on local generative model, that is ’handling occlusions for robust visual tracking’ and’robust tracking with a structured local model’.We propose an efficient method which incrementally updates an Eigen-basis and mean value as new observations arrives, used to learn the appearance of target during tracking progressing. To handle occlusions during tracking, we divide the samples into several same-size patches, judging whether one patch is occluded one by one. The occlusion matrix is used as dealing with each sample, to pick the non-occlusions patches for calculating likelihood probability. Experiments demonstrate that the effectiveness of our method when occluding occurs.This letter also presents a novel tracking algorithm that combines appearance feature and spatial information effectively. The geometric information is generated by transferring the detected SIFT points to a vector. Then the tracker combines’incremental visual tracking framework’ and occlusion detection to get a more robust representative model during tracking process.