Gait Recognition On Skeleton Extraction Based On Deep Learning

With the public’s emphasis on emerging biometric technology,gait recognition has become one of the hot spots studied by experts and scholars at home and abroad.Human gait has many natural advantages,such as long-distance recognition,difficult to hide and imitate,no contact with individuals,and uniqueness.Its recognition technology plays an important role in the field of public safety,but it still faces many difficulties that need to be solved urgently.Starting from the direction of deep learning,this paper studies the skeleton extraction algorithm by using a model-based method,and it solves the problem that the skeleton gait recognition task is susceptible to complex background interference.Using the characteristics of spatio-temporal correlation of gait information,this paper establishes a joint spatio-temporal network model to complete the final identification of gait.The main contents of this article are as follows:1)Improving skeleton extraction algorithm: In this paper,the Openpose extraction algorithm based on deep learning is used to extract human gait skeleton of the CASIA-B video database,thereby eliminating the complex background of gait video and avoiding the influence of covariates such as viewing angle and appearance.There are problems of wrong detection and missed detection when extracting pose skeletons by the traditional Openpose algorithm.In this paper,the Openpose algorithm is improved by using multi-scale and multi-scale,and bilinear interpolation is used to predict the skeleton joints of the current frame.Experimental results show that the improved algorithm corrects the problems existing in the extraction of gait skeleton,and obtains a more accurate and robust human gait skeleton.Finally,a gait cycle detection strategy is designed by using gait sequence periodicity,which extracts key frames and deletes redundant information,saving training time for subsequent accurate gait recognition.2)Establishing a joint spatio-temporal network model: Human gait data changes follow a powerful law of space-time,so this paper proposes a dual-stream 3D-Res Net＿AT-LSTM joint spatio-temporal network model to extract the gait invariant characteristics in the gait pose skeleton sequence and the gait RGB sequence.The first part of the network analyzes the advantages and disadvantages of the two-dimensional convolutional neural network and the three-dimensional convolutional neural network.Referring to the network structure of Res Net18 and Res Net50,a new three-dimensional residual convolutional neural network is designed to extract the spatial features of gait.The second part adds an encoding-decoding attention mechanism in the extended structure of the LSTM network to achieve the extraction of gait temporal features.The third part uses the score distribution fusion strategy to jointly obtain gait spatio-temporal features,and predicts the classification score to complete the final recognition of gait.3)Gait recognition experimental results and analysis: Based on the extraction of the skeleton of the CASIA-B dataset using the improved Openpose algorithm,this paper conducts experimental comparative analysis from different modules,different viewing angles,different covariates,network dimensions,and skeleton repair.Compared with the existing methods,the result shows that the dual-stream 3D-Res Net＿AT-LSTM joint spatio-temporal network designed in this paper has better recognition results,and the gait recognition on skeleton extraction based on deep learning is completed.