The Research On Key Techniques Of Multi-feature Fusion Based Human Tracking Across Camera Networks

With the technological development of computer vision theories and intelligent video analyses,human tracking across camera networks has become an important research topic in the intelligent video surveillance.However,human may vary greatly in appearance due to camera motion,target occlusion as well as the changes in target shape,pose,illumination and viewpoint from different cameras,therefore human tracking across camera networks is thus quite chalenging.To achieve robust human tracking across camera networks,the main research contents investigated in this thesis include DPM constrained multiple-kernel based human tracking,kernel based distance learning for person re-identification,and distance learning aggregation based person re-identification.To cope with camera motion and target occlusion during the human tracking so as to improve the robustness of human tracking,the human tracking based on deformable part model(DPM)constrained multiple-kernel,which efficiently integrates the DPM into multiple-kernel tracking is proposed in this paper.A DPM model of a tracked person is determined,and each part model of a DPM detected person is represented as a kernel.Each kernel is iteratively searched through the mean-shift algorithm based on spatially-weighted color histogram and histogram of oriented gradient(HOG),respectively,and the deformation cost provided by a DPM detector is used to constrain the movement of each kernel.Finally,multiple kernels are aggregated to determine the newly tracked position of the tracked person.The proposed human tracking based on DPM constrained multiple-kernel algorithm takes advantage of not only low computation owing to the kernel based tracking,but also robustness of the DPM detector,which can successfully track human more accurately under different moving camera scenarios.To deal with the changes of human appearance across different cameras so as to improve the accuracy of person re-identification,the kernel based distance learning is proposed for person re-identification.First,the kernel based local Fisher discriminant analysis(KLFDA)is proposed for person re-identification.More specifically,a feature fusion composed of color features from RGB,HSV and YUV color spaces,a texture feature caled uniform local binary patterns(LBP),and an edge feature caled scale-invariant feature transform(SIFT),is used to train KLFDA distance learning model.Second,the kernel based canonical correlation analysis(KCCA)is proposed for person re-identification.More specifically,an invariant feature called LOMO-FFN,which is a concatenation of local maximal occurrence(LOMO)and feature fusion net(FFN),is used to train KCCA distance learning model.The proposed two algorithms effectively improve the accuracy of person re-identification on some benchmark datasets.To further deal with the changes of human appearance across different cameras so as to improve the accuracy of person re-identification,the distance learning aggregation based person re-identification is proposed further.First,the simulated annealing based distance aggregation is proposed for person reidentification.More specifically,each optimized feature distance is obtained by the KISSME metric learning,including color histograms and salient color names based color descriptor(SCNCD)of multiple color spaces,and then individual corresponding feature distance weight in the distance aggregation is learned through a simulated annealing algorithm.Second,the distance aggregation of invariant features is proposed for person re-identification.More specifically,each optimized feature distance is obtained by two distance metric learning models,i.e.,cross-view quadratic discriminant analysis(XQDA)and large-scale similarity learning(LSSL),respectively,based on three invariant features,i.e.,LOMO,FFN,and LOMO-FFN,and then individual feature distance is fused through a min-max normalization.The proposed two algorithms clearly improve the accuracy of person re-identification on some benchmark datasets.