Research And Implementation On Dynamic Tow-view Motion Segmentation Based On Geometric Constraint

Motion segmentation for dynamic scene, as a vital research direction in computer vision, has a wide range of applications in image retrieval, image classification, motion analysis and interpretation of dynamic scene. The key issue of two-view motion segmentation problem is to estimate both numbers and parameters of motions simultaneously under the condition of non-parameter. This refers to the problem of fitting multiple motions to the scenes without knowing which feature points are moving according to the same model. The most challenging part is how to simultaneously determine the number of independent motions and identify the corresponding relationship between correspondence pairs and motion models with outliers'removal.At present, many approaches such as Expectation Maximization (EM) based algorithms, Factorization approaches and Generalized Principal Component Analysis (GPCA), etc, have been applied in the dynamic motion segmentation and have received satisfactory results. However, since the algorithms themselves need the number to be the input, they have to depend on the strong initialization of the number of motion models. Moreover, approaches of Factorization and GPCA which are algebraic methods still have some limitations and shortages of requiring a large number of feature points, being vulnerable to outliers and so on.In this thesis, design based on geometric constraint is proposed to deal with dynamic two-view scene motion segmentation problem. The research for iterative segmentation process of the design in this topic is a progressive process. For the limitations in previous methods, we improve the algorithm design from different perspectives step by step and first proposed 1) approach based on Dynamic Graph Sparsification (DGS). This method, a bottom-up process, regards motion segmentation problem as an abstract of geometric clustering problem. It implements clustering of feature points and scene segmentation through sparsifying process of geometric graph, which solves the problem of relying on the initial assumption of motion number and achieves non-parameter clustering analysis. Then we do research from the direction of density clustering and put forward 2) approach based on improved Gaussian Mean-Shift. Mean-Shift is one of the most sophisticated algorithms in clustering analysis of image density. Due to the limitation of Mean-Shift algorithm that it can be applied only to vector spaces, the method use the fitness vectors which are generated by scale and space transformation to convert the motion models into a high dimension vector space. The method obtains preferable stability by taking full advantage of the characteristics of Mean-Shift algorithm. Finally, by means of improvement with integration of the two methods proposed before, we present 3) approach based on divided-and-conquer and guided selection. This is a greedy-like method. Different from global clustering method and random sampling scheme of previous two methods, a guided selection is used to choose the most creditable hypothetical parameter as a seed motion model and only the motions similar to the seed will be selected as a sub-set, from which the iterative clustering process extracted the dominating motion. This approach can increase the time efficiency while ensuring accuracy.The approaches for dynamic two-view scene motion segmentation based on geometric constraint proposed in the thesis don't need any pre-knowledge of the number of motions and are non-parameter processes. There are no strict restrictions on the number of feature points and they can process a large number of motions without needing a lot of feature points. All approaches have less complexity for implementation.Two-view scene motion segmentation is an essential step of the initial processing towards interpretation of dynamic scene. It has a significant contribution and significance for the research in the fields of image retrieval, video analysis, target tracking and dynamic scene analysis, etc. This topic which can also be extended to non-rigid objects, three or more view motion segmentation and so on has broad prospects for development.