Research On Algorithms Of Video Object Tracking Based On Lie Group Manifold

Video object tracking technique is used more and more widely in military field such as missile dynamic measurement, unmanned aerial vehicles detection and missile guidance, and in civilian field such as intelligent surveillance, robot navigation and intelligent transportation, et al. It has become one of hot issues in the research fields of computer vision and pattern recognition.The core difficulties to be solved in the research work of video object tracking methods lie in both the geometric transformation of the object to be tracked, and the rapid changes of object feature data caused by the changes in the external environment such as illumination change and occlusion. It is a key technology to construct efficient and reliable models of geometric transformation and appearance feature for video object tracking. The existing video object tracking methods usually adopt affine transformation or projection transformation to describe geometric transformation of the object. However, the parameters in the models of affine transformation and projection transformation do not obey European vector space, but Lie group manifold space. Furthermore, the manifold data of image feature applied in the video object tracking method needs to be processed and analysed, which also obeys Lie group manifold space. So accurately analysing Lie group manifold structure of geometric transformation and appearance feature data of the object and building object tracking algorithms with high performance are of important academic research meaning and practical application value.In this dissertation, based on the analyses and conclusions of domestic and overseas researches, the two problems, geometric transformation and feature modeling of object to be tracked in video tracking algorithms under Lie group manifold theory, are investigated in depth. The major research contents and achievements are shown in the following areas:For tracking object with obvious geometric transformation, the geometric transformation model of the object is the key factor to guarantee the accuracy of tracking algorithms. Based on the analysis of the parameters of affine group and projection group(SL(3) group), and filtering algorithms, and the consideration that projection transformation reflects the process of objects imaging with more accurately, a novel object tracking method is proposed, by using particle filtering with dual manifold models on the projection transformation group. The method constructs dual manifold models, the covariance manifold used for the object observation model, and the geometric deformation on SL(3) group, adapted to utilize for object dynamic model. And dual manifold models combined projection transformation group with covariance matrix Riemannian manifolds, which can not only update the appearance model but also predict the dynamical manifold vectors so as to guarantee the accuracy of the tracking results. Extensive experiments prove that the proposed method can realize stable and accurate tracking of object with significant geometric deformation, and even with illumination changes and when an object is obscured the good tracking results can also be obtained.As to the problem that the feature model of the object is easy to be interfered by the noise under the complicated background, two methods for object feature modeling are proposed, and the corresponding object tracking methods are constructed. Firstly, based on partial least squares analysis is much robust and stable, a new object tracking algorithm based on partial least squares analysis using particle filtering with dual models is proposed. Using partial least squares analysis to describe the object region feature and applying affine transformation to describe the transformation process of the object, double dynamical particle filter models are built on Lie group and its tangent space respectively. Meanwhile, the proposed method constructs effective updating strategies of object feature space, which improves the accuracy of the tracking results when the object experiences appearance changes or when the background changes drastically. Experiment results show that the proposed algorithm can effectively filter out background noise and the tracking results are stable and accurate when the object is in complex background or it experiences temporary occlusion. Secondly, based on the analysis of anisotropic features with bilateral filtering, covariance object tracking algorithm based on bilateral filtering is proposed. Firstly, the image is dealt with bilateral filtering, and the gray and grads information of the filtered image is gained. Next, covariance matrix for the object to be tracked is constructed, and the tracking algorithm is designed, where the distance and similarity metric between covariance matrices adopts Log-Euclidean metric. Meanwhile, the rapid calculation of integral image is introduced, which effectively improves the computing speed of the covariance matrix. Experiments results show that the proposed method can track the object stably with illumination changes, and it can also track the infrared object accurately. Furthermore, it has the characteristics of high efficiency of calculation.On the research of the problem of multi-objects tracking, an algorithm for tracking multi-objects with mutual occlusion is proposed based on corner detection. Firstly, corner detection method based on bilateral structure under Lie group is constructed. Secondly, K-NN algorithm is adopted for labeling the corners in the occlusion region, and comer matching method is designed which can distinguish each tracking object. Thus, the objects with mutual occlusion are separated successfully. Finally, a multi-objects tracking algorithm is proposed, which can effectively distinguish the objects with mutual occlusion. Extensive experiments show that the proposed algorithm can effectively realize the stable multi-objects tracking when mutual occlusion occurs.The theory of Lie group manifold is introduced to the field of video object tracking in this dissertation, and the problems of geometric transformation modeling and feature modeling are researched systematically. Furthermore, simulation and experimental verification are done for the proposed algorithms, the experimental results show that the proposed algorithms can effectively overcome the influences caused by the object deformation, occlusion, complex background, illumination changes and have been verified to be feasible, available and advanced.