Visual Object Tracking Via Sparse Learning

As a linkage between the information processing technology located at the bottom of computer vision system and the high-level content analysis,researches on visual object tracking have been of great significance in the field of intelligent auxiliary driving and video surveillance.By searching for the exact location of tracking target in the next frame and feeding it back to driver assistant system or video monitoring system for tracking,the crucial motion information can be provided by visual tracking algorithms for video sequence analysis and understanding.This dissertation presents an in-depth study centering on sparse learning based trackers,aiming at how to design a robust visual tracker under complex environment and how to balance tracking efficiency and robustness in practical applications.The sparse learning theories are overviewed in Chapter 2,as well as the fundamentals of sparse learning based tracking algorithms.A RGBD and sparse learning-based tracker is proposed in Chapter 3 on the basis of rationales introduced in previous chapter,which significantly improves tracking robustness under drastic illumination changes and various occlusion conditions via efficiently exploiting the valuable information provided by range data.A Binocular Consistent Sparse learning based Tracker(BCST)is proposed in Chapter 4,which further improves tracking robustness under drastic posture changes and tracking efficiency via efficiently employing the appearance and depth information from the binocular configuration.To satisfy the real-time demands in practical applications,A Hierarchical Convolutional Features and Sparse learning based Tracker(HCFST)is proposed in Chapter 5,which simultaneously optimizes tracking accuracy and efficiency by pruning invalid convolutional features most of whom are not related to the tracking target.Besides,a particle filter based hierarchical target localization method is also given to efficiently deal with the scale adaptability problem.The main content and innovation of this paper are as follows:1.A novel RGBD and sparse learning-based tracker is proposed,which makes the best of valuable information provided by the range data and significantly improves tracking performance under drastic illumination changes and various occlusion conditions.Based on the sparse learning framework,the range data is firstly exploited for occlusion detection within the target region,then the detected area is formulated as a model to generate the occlusion templates.Integrating these occlusion templates with the existing over-complete dictionary endows the proposed tracker with the ability to handle various extreme occlusion conditions.Considering range data is insensitive to varying illumination,an extra depth view is integrated to replenish color image-based visual features for robust appearance modeling.Finally,an occlusion detection method employing depth based histogram analysis is proposed to efficiently determine the proper time for the template update.Extensive experiments on both KITTI and Princeton data sets demonstrate that the proposed tracker outperforms the state-of-the-art tracking algorithms,including both sparse and RGBD-based methods.2.A novel Binocular Consistent Sparse learning based Tracker(BCST)is proposed,which makes the best of the appearance and depth information from the binocular configuration.BCST further improves tracking performance under drastic posture changes,as well as tracking efficiency.Firstly,valuable prior appearance of tacking object obtained through the second camera is integrated into an augmented dictionary via the proposed crossover templates,thus tracking accuracy and robustness under sharply changed posture can be improved.Then a special depth consistency constraint is designed to replenish row sparse constraint in the joint sparse learning,thereby making the proposed tracker more stable.At last most of the stray particles can be removed according to the depth consistency property with the assumption of small range variations of tracking object between frames.This property facilitates more efficient tracking results.Qualitative and quantitative evaluations on KITTI Vision Benchmark show that the proposed BCST demonstrates superior performance over the state-of-the-art tracking algorithms,including both the sparse and stereo-based methods.3.A novel Hierarchical Convolutional Features and Sparse learning based Tracker(HCFST)is proposed,which exploits features extracted from convolutional neural networks and models the target appearance with efficient correlation filters.HCFST is designed for real-time tracking of moving targets under complex environment.A sparse learning based screening process is proposed to prune invalid convolutional features most of whom are not related to the tracking target,thereby significantly improving tracking performance and efficiency.Meanwhile,a particle filter based target localization method is designed.By hierarchically inferring the optimal position of tracking target from the top layer to low layer,the problem of scale adaptability can be solved efficiently.