Cross-covariant Gait Recognition Based On Multi-feature Fusion

With the continuous development of computer vision field,gait recognition,which is an identification method based on pedestrian gait,has gradually attracted scholars’ attention.Compared with other identification technologies,gait recognition can realize long-distance,non-contact identification,and the recognition process does not require the cooperation of the subject.The existing gait recognition methods are less robust in the case of changes in covariate variables such as pedestrian’s clothing and carrying objects,and do not fully consider the differences in the motion rules of different parts of the human body during walking,and lack comprehensive consideration of motion and spatial-temporal gait feature representation.Based on deep learning and computer vision,this thesis focuses on feature extraction and feature fusion of motion feature and spatial-temporal feature in gait recognition.The specific research work is as follows:(1)Motion gait feature extraction.Motion gait features include pedestrian motion change information,which improves the robustness of gait features under cross-covariant conditions.Therefore,this thesis proposes motion excitation and embedding self-attention modules,which are respectively used to extract temporal motion change feature and enhance the expression of motion law.Under three different appearance conditions in the CASIA-B dataset,the proposed modules improve the network performance by 0.1%,0.5%,1.3% and 0.2%,0.1%,0.6%respectively compared with the baseline.(2)Spatial-temporal gait feature extraction.The different horizontal positions of pedestrians have different movement rules.Mining the different movement rules of local parts helps to accurately learn the pedestrian movement pattern.Therefore,this thesis proposes a nonlinear horizontal mapping module,which provides different weights based on horizontal local positions;Based on GLFE module,this thesis proposes an improved FFE module to extract spatial-temporal gait features for different horizontal parts.Compared with the original network model,the nonlinear horizontal mapping module in the CASIA-B dataset has improved by 0.1% and 0.4%respectively under NM and CL conditions,and the FFE module has improved the model by 0.5%,0.5% and 1.0% respectively under NM,BG and CL conditions.(3)Gait feature fusion.The relevant experiments in the field of action recognition have confirmed that motion features and spatial-temporal features are complementary,and feature fusion can improve the recognition performance of the model.This thesis proposes the motion branch,which can focus on refining the information of motion change features in the horizontal and temporal dimensions,and uses feature splicing to fuse the motion and spatial-temporal gait features.The motion branch and feature fusion proposed in this thesis bring 0.5% and 0.3% improvement to the improved network model under NM and BG on CASIA-B dataset.After integrating the research contents of each part,the complete network model proposed in this thesis achieves the state-of-the-art recognition accuracy of 98.4%,95.5% and86.0% respectively under the NM,BG and CL conditions on the CASIA-B dataset.