Research On Visual Tracking Method Under Complex Environment

Visual tracking is a very challenging task in computer vision community, because the object appearance exists many complex changes in realistic scenarios, such as scale changes, partial occlusion,3D rotation, illumination changes, object deformation, and so on. To address the problem, this thesis gives a deep survey and study of classic visual tracking algorithm based on discriminative method, and proposes several efficient visual tracking methods. The major work and contributions lie in several fields as follows.1. A novel leaning to rank algorithm, referred to as Ranking Vector SVM (RV-SVM), introduced to the visual tracking community, and propose an efficient visual tracking algorithm based on RV-SVM. First, the algorithm employs a sparse random measurement matrix to extract the multi-scale image features of samples. Then, a Median-Flow tracker is used to estimate the location of target object roughly, and construct training data sets. Finally, train the RV-SVM classifier online, and distinguish the object from background. The proposed algorithm can deal with many appearance changes in realistic scenarios, such as scale changes, partial occlusion, illumination changes,3D rotation and fast object motion.2. A spatial divide and conquer approach is used in circulant matrix tracking algorithm. The proposed algorithm subdivide object into smaller regions, then, tracking each sub-target using the circulant matrix tracking algorithm. For the tracking results of each sub-target, compute the confidence values. At last, update the location of target object according to the tracking results of sub-target and confidence values. The proposed algorithm takes advantage of the spatial distribution information of the target. Compared with the circulant matrix tracking algorithm, the proposed algorithm can track stably under complicated environment, such as fast motion, scale variation and severe occlusion.3. A novel scale adaptive object tracking algorithm based on log-likelihood image is proposed. First of all, according to the color difference between target and background, the log-likelihood image is built, and the mathematical morphology method is adopted to process the log-likelihood image. Then, a new ellipse fitting method is applied to the log-likelihood image to estimate the scale and orientation changes of the target. Finally, update the bandwidth of Mean-Shift kernel function, and tracking the object by Mean-Shift iteration. The proposed algorithm is simple and effective, can well solve the problems that estimate the scale and orientation changes in object tracking.4. Based on the foregoing studies, a robust and efficient visual tracking demonstration system is developed. The system supports three video input ways, namely, local video files, webcams and image sequences. What’s more, five visual tracking methods are integrated in the demonstration system. In addition, the system provides many useful functions, such as video open, play, pause, and visual tracking, algorithm switch, pause tracking.