Research On Tracking Algorithm For Visual Target Based On Statistical Learning

Visual object tracking is a key problem in computer vision, it has a wide range of applications in military andurban fields. Although a huge of effective tracking algorithms has been proposed, there are still lots of difficultiessuch as illumination changes, pose variations, deformations, occlusion, clutters. Therefore, it is a very challengingtask to design an accurate and robust algorithm for visual object tracking.In this dissertation, the research is focused on the twofold in visual object tracking in the framework of particlefilter, appearance representation design and observation likelihood design. The main contributions of this dissertationare summarized as follows:1. A novel algorithm based on nonnegative least square estimation for visual object tracking has beenproposed. The algorithm offers a new perspective for template matching which only consider each candidateseparately. The template model is approximated by the linear combination of candidates in the space spanned by thefeature vectors associated with particles. The coefficient of the linear combination is constrained as nonnegativity andis obtained through solving a nonnegative LS problem, and then it is used to evaluate the similarity between thetemplate and candidates. Experimental results demonstrate that the proposed algorithm has better tracking accuracythan the compared algorithms and strong robustness against noise.2. A novel algorithm based on the local-constrained least square estimation for the visual object tracking hasbeen proposed. In this method, the relationships among these local patches of the object are exploited. Thetopological relation can reflect the object’s structure which is helpful for handling occlusion and similar objectdistraction problems. Then weights associated to particles are measured based on the Bhattacharyya coefficient.Experimental results demonstrate that the proposed algorithm has better tracking accuracy than the comparedalgorithm, and can track the object scale well. 3. A new tracking algorithm based onl2regularized least square optimization has been proposed. Templateand candidate observation image can be linearly represented on the predefined dictionary thoughl2regularizedleast square estimation. The estimation has analytical solution, which make the coefficient computation very fast. Themodeling for latter incorporates the prior information, in order to increase the robustness to pose variance,illumination variance, clutter, occlusions and so on. The trivial template set is introduced into the dictionary to explainthe noise. The tracking task is implemented based on particle filter method and the observation likelihood is definedas the Euclidean distance for its sparse representation to the prior model. The similarity measure has furtherstrengthened the algorithm’s discriminative power.4. A new visual tracking algorithm based on multi-task sparsity subspace learning and incrementalGrassmann update has been proposed. This algorithm gives multi-task sparsity learning model based on subspace,which can effectively handle the background noise or outliers to the object appearance representation. The trackingproblem is performed in the geometrical particle filter on affine group Aff (2). The state model is described by afirst-order AR process. This provides an effective search mechanism. The observation likelihood is based on thel1norm about reconstruction residual. In order to tackle occlusion issue, the dual vector is introduced into the likelihood.The subspace update can be considered as an optimization problem on Grassmann manifold. In gradient descentmethod, the step size along the geodesic is very important, thus, an adaptive step size strategy is given in this study.The proposed subspace update strategy can be able to adapt to slow or abrupt appearance changes, which effectivelyavoids the model drifting problem and is obvious to increase the robustness and accuracy of the tracking algorithm.