Visual Tracking Via Local And Nonlocal Similarity Learning

Visual tracking is a fundamental problem in computer vision with a wide range of applications in human-computer interfaces, surveillance,traffic control,motion analysis. Although much progress has been made in the past decades,it remains a challenging task due to some factors. The main work of this thesis is researching visual tracking algorithm via local and non-local similarity learning. The main contributions of our work are summarized below:(1) We present a patch based tracker which adaptively integrates the kernel correlation filters with multiple effective features. To take advantage of the useful information from different parts of the target, we train each template patch by kernel correlation filtering method, and adaptively set the weight of each patch for each particle in a particle filtering framework. Experiments illustrate that this scheme can effectively handle the occlusion problem. Moreover, the effective features including the HOG features and color name features are effectively integrated to learn the correlation between the target and background, the candidate patches and template ones, which further boosts the overall performance. Extensive experimental results on the benchmark demonstrate the proposed approach performs favorably against some representative state-of-the-art tracking algorithms.(2) We propose a visual tracking via non-local similarity learning method.Either global or local feature representations have been widely exploited for visual tracking. However, most of these representations describe a target appearance with a fixed spatial grid layout without considering the interaction between different grids,and hence may adversely affect their performance when the target appearance suffers from large-scale pose variations. In this thesis,we learn a similarity function that considers the interactions of features in the grids not only from the same spatial positions, but also from different positions, thereby taking charge of the nonlocal information of the target appearance. Specifically, we explore the polynomial kernel feature map to characterize the nonlocal similarity information of all pairs of grids among the target and its background samples, and combine these feature maps as the target representations. Moreover, we learn a linear logistic regression classifier with online update and integrate this classifier into a particle filtering tracking framework.Extensive experimental results on the benchmark demonstrate the proposed approach performs favorably against some representative tracking algorithms.