| In the intelligent video surveillance network,an important development direction is the pedestrian relay tracking in multi-cameras,which has a high research value and a pressing application need.Revolving around the topic in this dissertation,relatively systematic reviews on the present research are given.Careful studies on some important theories and methods are made.Based on the study,new attempts are made for improvements,resulting some better performances in experiments.1.A comparative and classification study is made on the present target detection methods.Deeper studies are made on the currently common methods such as background subtraction method and Gaussian mixture model method.For the difficulty of foreground segmentation when the target and background are too close in grayscale,a novel method is proposed using gray contrast of a pixel with its neighbor pixels as the second clue.The contrast deviation of the current frame off the background model at a pixel is evaluated with the Gaussian noise model.The occurence probability of the deviation is used to adjust the target detection threshold at the pixel,helping improve the target segmentation performance of the traditional background subtraction method and Gaussian mixture model method.As the proposed method is not based on the richness of local texture,the improvement effects can also be seen in the flat area of the image,and the robustness of texture in the non-flat area is maintained.2.A comparative and classification introduction is given on the present methods for single camera tracking and the relevant target features.An in-depth study is made on the particle filter theory and the implementation of the method.As the target tracking process of single camera may be influenced by the shadow casting,a color feature of RGB chroma space is proposed to enhance the robustness to brightness variation.3.A summarizing study is made on the basic idea and related problems for target handing over across cameras with overlapping field of view.An introduction is given on the common methods for target handing over.Deeper studies are made on the geometric constraint processing.For the stability of target feature across cameras,a propose is given using the constriant condition that the target has the same reflective appearance in the overlapping FOV(field of view),leading to an invariant color feature independent of cameras.Further more,a bag of texture words model is proposed.Several regions on the target surface are found via cluster analysis where U-LBP distribution is relatively stable.The centroids of clusters are defined as different words.Then,the description of the target across cameras can be given by the occurrence frequency of the words on its surface.For better use of the proposed features and based on the Gaussian error distribution hypothesis,a Bayesian belief network implementation method is proposed,which makes multi-feature fusion of coordinate,color and texture for re-identifying targets across cameras.Based on the image block synchronization,an automated generation algorithm for camera FOV line is proposed.4.A summarizing study is made on the basic idea and common features for target handing over across cameras with no overlapping field of view.A classification of common methods in the color deviation correction is given.In order to get good stability of target feature across scenes,an invariant color feature is proposed.With the reflective imaging model of target and photoelectric response model of camera sensor,the target pixel color is decomposed into brightness item,tint item and the target inherent item.Then,by re-construction,the invariant color feature is obtained which is rubust to lighting intensity,lighting temperature,camera parameters and some viewpoint change.Further more,with the use of Fourier transform which can separate the surface structural information from the overal brightness as AC and DC signals,some useful attempts are also made on targets re-identification across scenes in frequency domain. |
PDF Full Text Request