Salient Regions And Topic Model Based Proto-objects Detection And Visual Target Tracking

With the development of computer vision technology and expanding demand for practical application, moving target visual tracking technology has become the focus of domestic and foreign researcher because of its applications in areas such as safety monitoring, robot navigation, intelligent transportation and unmanned potential. Although research on visual tracking has lasted for decades, various noisy factors including background clutter, occlusion, distraction, and illumination variance still cause problems in practice, which makes visual tracking a challenging task.In order to improve the robustness of visual tracking, reduce the tracking error rate, we propose a biologically inspired framework of visual tracking based on proto-objects. The proposed approach simultaneously estimates the states of the target and proto-objects over time, so it achieves superior performance. In this thesis, the main contents and results are as follows:Firstly, given an image sequence, saliency detection for the image is operated on. Usually the target is more likely to appear in the places with higher saliency values. The method of saliency detection in this thesis is between the low-level feature based representation and the object-level semantic representation. The saliency map combines the bottom-up mechanism with the top-down mechanism, so that the extraction is closer to human vision system feature.Secondly, we use salient regions and topic model based proto-objects detection. We select Normalized Cut after comparing various segmentation method and then object classes on each segment region are estimated by using probabilistic Latent Semantic Analysis(PLSA), we obtain the final proto-objects after the topic being Revealed.Later, In order to avoid the condition that most of the segmentation based detection methods have relied too much on the quality of image segments, a new method is proposed based on Multi-Level-probabilistic Latent Semantic Analysis(ML-PLSA) algorithm. The detection results are obtained by fusing estimation results at multiple levels. So it performs better than traditional methods in both accuracy and robustness.Finally, the target is tracked with Bayesian approach based on spatial information of the proto-objects. States of the target and proto-objects are jointly estimated over time, so that their correlations to the target are much more stable during the tracking process.In the paper, Gibbs sampling has been used to optimize the tracking method. The EM algorithm is utilized to reveal the topics and optimize the topic assignment through maximum log-likelihood, and Markov network can infer the states of both the target and the proto-objects during the tracking process. Experimental results demonstrate that the proposed approach achieves good performance which can robustly deal with occlusion, distraction, as well as illumination variation. Experimental results also demonstrate that the proposed method outperforms the state-of-the-art methods in challenging tracking tasks.