Studies On Moving Object Segmentation In Video Sequences And Recognition

The purpose of this thesis is to segment moving object in stationary background ,then recognize it.In the moving object segmentation part, we proposed a method extracting moving object by utilizing the color information along with the moving information.Color image segmentation primarily uses a non-parametric clustering method based on partial space information intervening iterative mean shift procedure.By setting only a few parameters,we can subsume each pixel of the image into its corresponding density pattern and realize clustering. In this thesis,a way of finding connected areas with identical marks was introduced to divide the image into different independent areas.Owing to the fact that the HOS (Higher-order statistics) have a property of eliminating Gauss type noises,it was used for motion detection.Canny operator was used for extracting the contour of the motion area.After filling the inner of that motion area,rough motion template can be obtained. In such a template,the ratio of regions with identical marks to the corresponding independent region in the original image can be computed.If the ratio is larger than a predetermined threshold,the corresponding independent region is regarded as a region to be merged,otherwise,the corresponding region does not belong to the moving object which will not be merged.Finally,the moving object can be extracted by merging all the predetermined regions to be merged.In the recognition part of this thesis,all discussions are constrained in two dimensions. In order to eliminate the effect of the translation,scaling,skewing and rotation on the recognition result.We proposed a novel method.In this method,the covariance. matrix of the image under a given pattern are firstly computed.Rotating the imageaccording to the eigenvectors of the covariance matrix,moving object template irrelevant to the new coordinate system can be obtained. Adjusting the height and width of the object template according to the eigenvalues of covariance matrix,so a translation ,scaling,skewing normalized form of the moving object can be obtained.The features of the moving object are constructed by using the 8-Zernike moments of the contour of the prementioned normalized form.Before recognition,some samples of known classes are used to determine the centers of all the classes. Moving objects were classified by evaluating the Euclide distance between the features and those centers determined before.Simulation results verified that the method proposed in this thesis is well validate for recognizing objects with obvious differences.