Online Latent SVM For Scale Adaptive Visual Tracking

Visual tracking is a hot research topic in computer vision that finds applications in many fields, such as human machine interaction, video surveillance, advanced driver assistance systems, etc. Tracking by detection is a popular framework in visual tracking which takes advantage of various machine learning techniques to train a classifier that can separate the object from background with training examples. The learned classifier is used to detect the image patch at possible locations from the frame to be tracked and to find the location with maximum possibility. With the discussion of traditional tracking by detection algorithms, the dissertation points out the problems exist in these algorithms and proposes the corresponding solutions.In the tracking of general object, the appearance of the object may vary and the classifier for the object cannot be trained before tracking. Tracking by detection algorithms make use of frames obtained during tracking to update the classifier online. Most of these algorithms train a new classifier with examples in the new frame, and apply the online weighted sum to the parameter of the new classifier and old classifier in order to form the current classifier. This approach makes it possible for the classifier to track the new appearance of the object with one drawback that the classifier forgets the appearance of the object in history frames after several iterations. To solve the problem, this dissertation proposes using online support vector machine(SVM) to learn the appearance of the object in key frames. Online SVM can select key frames adaptively with the appearance of the object in every frame and the status of the SVM, and the weight of every frame can also be obtained with the optimization algorithm which is different from the update strategy of traditional online tracking algorithms that learn new classifiers for all frames. Compared with traditional online tracking algorithms, online SVM can learn representative object appearance during tracking, which makes it possible for the classifier to handle the problem of varied object appearance. The max margin criterion in SVM also makes the algorithm based on online SVM very robust.Most tracking by detection visual tracking algorithms do not consider the variation of object scale during tracking, or search all possible scales in the frame by brute force to find the window contains the object. However, there exist two problems in brute force searching. Firstly, the quantity of searching windows is n times larger than the original one, where n is the number of searching windows with different scales. In occasions where the scale of the object varies dramatically, large quantity of windows with different scales is needed for the searching, which would significantly decrease the efficiency of the classifier. Secondly, the features of the image patches in windows with different sizes may be similar, especially when the texture of the object is not obvious. It is clear that this would affect the output of the classifier and make the scale of the tracking window unstable. To handle these problems, this dissertation proposes to treat the scale as a latent variable and take advantage of latent SVM for scale update. In latent SVM, the learned scale in previous frame is used to search the location of the object in current frame. Once the best location is obtained, multiple scale searching is applied in the best location to find the best scale. The obtained best location and scale are used to sample training examples for the training of the SVM, after which the SVM is used again to find the best scale. The searching of scale and training of the SVM are processed iteratively until convergent. The latent SVM transfers the scale searching from detection phrase to learning phrase, which guarantees the efficiency in the detection phrase and decreases the possibility of drifting.Finally, this dissertation combines online SVM and latent variable and proposes the scale adaptive visual tracking algorithm based on online latent SVM. The proposed algorithm can learn the appearance of the object in key frames and learn the scale of the object in the training phrase of the SVM. Experiments compare the algorithm proposed in this dissertation and several other popular algorithms. Experimental results show that the accuracy of the proposed algorithm is higher than other algorithms on the test video sequences.