Gait Analysis Based On Human Skeleton From Multi-view

In recent years,as a new biometrics technology,gait recognition has attracted more and more attention of researchers.Long-distance,difficult to disguise and non-contact are the advantages of gait recognition technology that distinguishes it from other biometric technologies.The randomness of human movement makes it difficult to determine the walking direction,and the gait features of human change with the change of the view.Therefore,it is necessary to analyze human gait from multi-view.Around this topic,gait analysis is carried out by a human skeleton that is less affected by the view.The main research in this thesis is as follows:1.Human motion region segmentation.In order to solve the influence of dynamic background and human shadow on human region segmentation,this thesis makes qualitative and quantitative analysis of color channel.Finally,this thesis chooses color channel S,V and Cr to form a mixed color space.In this mixed color space,an improved Codebook algorithm is used to segment the foreground of multi-view video sequence.Experiments show that the proposed algorithm reduces the influence of dynamic background and human shadow on human region segmentation,quickly extracts human moving objects under multi-view and improves the robustness.The proposed algorithm improves the detection accuracy by 12.64%compared with single color space.2.Human skeleton extraction and transformation from multi-view.This thesis analyse different algorithms for extracting human skeleton and improving ZS(Zhang and Suen)thinning algorithm to solve the problem about loss of skeleton information.Experiments show that the improved ZS thinning algorithm can quickly extract the human skeleton with strong topology and connectivity.The accuracy of extracting skeleton is 12.64%higher than ZS thinning algorithm.For human skeleton from different view,gait trajectories are mapped to side view by establishing View Transformation Model.The experimental results show that the gait similarity between the non-side view trajectory and the side view trajectory is about 4 pixels on average.3.Analysis and classification of human gait characteristics.By analyzing various gait features,the gait features with large discrimination and adapting to the viewing angle are extracted to form the feature vector used as the classification feature.The extracted feature vector is used as the classification standard,and the Support Vector Machine(SVM)is used to classify different gait sequences of the same person,different view of the same person and different human gait sequences to verify the validity of the feature vector.Experiments show that the feature vector has higher discrimination and can adapt to changes of view.Experiments show that the proposed algorithms can improve the accuracy of multi-view video sequences in CASIA-DatesetB database by at least 10%.The experimental results show that the final gait feature vector is less affected by the view and has higher discrimination.