Research On Key Techniques Of Vision-based Moving Objet Tracking

Visual object tracking has been widely applied in the fields of intelligent video surveillance,3-D reconstruction,robot technology,etc.However,due to the complexity and the diversity of the video image data,and the uncertainty character in the moving object itself,designing a visual target object tracking system with high robustness,better real-time and high accuracy in variety of complex scenes is still a hot spot and difficulty of visual object tracking study.In this dissertation,guided by the generated model and the discriminative model,we detailed analysis and make an intensive study of appearance representation design,search strategy,similarity measurements and template updating strategy in visual tracking.Then some new ideas and methods,which combine international academic development and demands for practical application,are proposed to solve the difficult problems of visual object tracking.The main contributions of this dissertation are summarized as follows.Firstly,a novel memory gradient pursuit algorithm(MGP)for compressive sensing signal reconstruction is proposed.The algorithm adopts regularization orthogonal matching strategy to select atom sets fast and efficiently,then the search step size is determined by nonmonotonic linear search strategy and update direction is fixed with the memory gradient algorithm.The proposed algorithm is a novel directional pursuit method.It takes fully advantages of globally fast and stable convergence of memory gradient algorithm with Armijo line search to avoid local optimal solution.After that,sparse coefficients are achieved.Experimental results demonstrate that,the algorithm has reduced time for signal reconstruction while keeping the accuracy.Additionally,the proposed algorithm outperforms other reconstruction algorithms in accuracy when some conditions are satisfied.Secondly,an optimized sparse representation tracking algorithm based on memory gradient pursuit algorithm is proposed.The algorithm provides an efficient appearance model which is fit for sparse representation scheme,and fast solving method for L1 norm minimization problem.A non-overlapping covariance descriptor is introduced into the appearance model.Then the new appearance representation takes fully advantages of the block information which is helpful for handling occlusion and background clutters problems.Moreover,an adaptive scaled unscented transform method with lower computation cost is adopted to approximate the covariance matrix.Then,the similarity metric of covariance descriptor is transferred from manifold to Euclidean space.The algorithm takes advantages of fast and stable convergence of memory gradient algorithm to reduce the reconstruction time of sparse coefficient.On the basis of the descriptors with templates updated strategy,the robustness to complex scenes and model drift is improved.Experimental results demonstrate that the proposed algorithm has better tracking accuracy than the compared algorithm,and is robust to the scale changes.Thirdly,a novel fast memory gradient algorithm for covariance tracking based on JBLD(Jensen-Bregman LogDet)metric is proposed.The JBLD metric with low time complexity is utilized to measure the similarity of covariance features,which can reduce the computational burden of the similarity metric especially for large symmetric positive definite matrices under Riemannian space.Moreover,the JBLD metric contributes to fast computation of the gradient under the framework of the gradient based optimization algorithm.By minimizing the metric function based on memory gradient method,the proposed algorithm can search the best matched candidate efficiently.We test the proposed tracking method on test baseline dataset.Both quantitative and qualitative results demonstrate the effectiveness of the proposed algorithm compared with other state-of-the-art methods.Finally,an online discriminative feature selection(ODFS)real-time object tracking algorithm based on distribution fields descriptors(DFs)at instance level is proposed.The algorithm adopts layers of DFs feature based on online discriminative instance feature selection algorithm to represent the target,which is more efficient for the appearance model.Then,replacing DFs with channel representation based approach helps to reduce the computation time greatly.The most discriminative but fewer features are selected by ODFS,which can enhance the distinguishment ability and real-time of the algorithm.Therefore,the proposed algorithm can effectively avoid the model drifting problem.The experiments on real scenarios demonstrate that the proposed algorithm can effectively track object in complex circumstance.