Visual object tracking technology, which embraces computer graphics, pattern recog-nition, matrix theory and probability theory, is one of the popular research areas. Withthe development of object tracking technology and robotics, visual tracking is widelyused in the field of robot control. However, since the visual servo is generally used in thecomplex environment of production, where exists uncertainties of illumination change,appearance variation, target occlusion, and complicated background, it is required thatthe object tracking algorithm has high robustness, accuracy and instantaneity. As a re-sult, the present paper probes into the object tracking algorithm with the purpose ofenhancing its robustness, accuracy and instantaneity in complex environment. The mainresearch questions of this paper are as follows:Firstly, Gabor features are introduced into sparse representation target trackingalgorithm On the basis of analyzing the basic theories of object tracking, a sparse repre-sentation tracking algorithm which uses Gabor features is proposed. The basic theoriesinclude particle filter theory, sparse representation theory, affine transformation theory,and so on. In order to adapt to changes of illumination and appearance, Gabor featuresare introduced into sparse representation object tracking algorithm. In order to adapt tovariations of targets’ appearances, the template dictionary is updated by using the PCA(Principal Component Analysis). The information of tracking results is merged into thetemplate dictionary whose dimension is fixed, thus making the tracker more robust.Secondly, a target-background weight block sparse representation algorithm is pro-posed. In order to improve the instantaneity of object tracking algorithm, a kind of vari-ance particle filter is applied. This method filters out abundant particles in the particlefilter framework and reduces the amount of calculation by the object tracking algorithm.In order to further improve the robustness and instantaneity of target tracking algorithm,a kind of target-background collaborative sparse representation method is proposed. Themethod fuses background information into spares representation template dictionary, sothat improves the ability of detecting target and background. Meanwhile, overcomes thelarge calculation of solving the1-regularized least square equation. In order to adapt tothe occlusion of target, a weight block sparse representation algorithm is proposed. Thealgorithm blocks sample into several pitches, each of which obtain a weight by iteration.When the blocks are occluded or disturbed by noises, they will get a low weight to limitits impact on the overall results, so that the algorithm is more adaptive to occlusion.Thirdly, the structured sparse representation tracking algorithm based on the rank- ing of coefficient and residual error is explored. In order to enhance the accuracy ofobject tracking algorithm, the theory of structured sparse representation is introduced.This theory not only contains overall information of the samples, but also includes spaceand local information of the samples. In order to make coefficients of structured sparserepresentation more effective, a kind of sparse coefficient score is designed to measure ofthe similarity of samples and the target. In order to make residual errors of structuredsparse representation more effective, a kind of residual error score is designed to measureof the similarity of samples and the target. In order to merge information of coefficientand residual error, whose properties are different, a ranking and information entropybased information fusion method is proposed. The method is able to fuse two kinds ofscore effectively, and make the tracking algorithm more robust. The template dictionaryis renewed by a method that combines sparse representation and PCA together, thusenabling the tracker to adapt to the variation of target appearance and background.Fourthly, under the framework of structured sparse representation,the present the-sis analyzes the object tracking and detecting algorithm on the basis of learning andclassification. In order to improve the robustness of the object tracking algorithm, Bayesclassifier is proposed. This classifier takes target tracking as a2-value classification, excelsat detective job and requires a small amount of calculation. The target tracking algo-rithm fails to work in the circumstance of target disappearance, while combining targetdetecting algorithm and object tracking algorithm could solve the problem. The detect-ing algorithm is able to recognize appearance and disappearance of the target throughscanning.Finally, in order to demonstrate the performance of target tracking algorithm usedon motion control, a kind of object tracking motion control platform is designed. Similarto visual servo, this platform enable the cameras to trace a moving object. It is composedof computers, motion control cards, servo drivers, motors, x-y motion platform and cam-eras. Vision information, which is collected by the cameras, is send to the computer forimage procession. The target’s state, which is calculated by tracking algorithm, is usedto control x-y motion platform. The target remains in the center of the images because ofthe moving of the x-y motion platform. Two kings of experiments are designed to demon-strate performances of the object tracing motion control platform. They are qualitativeand quantitative experiments, respectively. The qualitative experiments demonstrate thefeasibility and rationality of the platform. The quantitative experiments demonstrate theperformance of tracking algorithm on the motion control through test data.In general, the present thesis probes into the sparse representation object track- ing algorithm under the requirements from visual servo, with the purpose of enhancingrobustness, accuracy and instantaneity of the object tracking algorithm. |