Gait Recognition Of Feature Decomposition Based On Appearance And Skeleton

Gait Recognition is a means of identifying the target.It is not as accurate as Face Recognition,Iris Recognition and other biotechnological recognition methods,but it does not need the cooperation of the target,and can also accurately recognize at a long distance.Therefore,it is of great significance to improve the accuracy of Gait Recognition and apply it to Person Re-identification,HumanComputer Interaction,Cameras and Surveillance and other fields.Because the accuracy of gait recognition is easily affected by clothing,shooting angle,speed,light intensity and other factors,researchers’ work mainly focuses on better extracting more effective gait features unrelated to these factors.At present,in the field of Gait Recognition,the method of Deep Learning is mostly used to build models.According to different input types,Gait Recognition Algorithms can be divided into two categories: Appearance-based Algorithms and Skeleton-based Algorithms.The main content of this thesis is to study the accuracy of Gait Recognition Algorithm under different perspectives according to different input types.Assuming that the human gait feature is a whole,it can be decomposed into various sub-features of different categories.Based on this idea,this thesis proposes two gait recognition models according to different input characteristics.Firstly,an Appearance-based Gait Recognition Algorithm is proposed: Taking a sequence of gait silhouettes as input,the algorithm decomposes gait features into the features in the 0° and 90° viewing directions.An Orthogonal-view Feature Decomposition Network based on GaitSet(OFD-GaitSet)algorithm is proposed;Secondly,a Skeleton-based Gait Recognition Algorithm is proposed: Taking skeleton keypoints as input,the algorithm extracts skeleton keypoints from the gait silhouette sequence and performs block processing.A Gait Recognition Algorithm based on skeleton blocking-features is proposed,which is called PartGraph.The specific research content is as follows:(1)Using a sequence of gait silhouettes as input,the gait information is decomposed into a subset of gait information with two views: 0° and 90°.And the OFD-GaitSet algorithm is proposed.Firstly,the algorithm improves the setting of the gait gallery so that each sample in the gallery contains gait information with two views: 0° and 90°;Secondly,the algorithm designs two Feature Extraction Networks,which extract the gait sub-features of the gait silhouettes sequence from two views.At the same time,the View Identification Network and Distance Block are used to weight the Euclidean Distance between the gait sub-features and the gallery’s,and the recognition results are obtained through comparison.This algorithm uses Cross Entropy Loss and improved Triplet Loss for training.Experiments on the CASIA-B dataset show that the average Raank-1 accuracy reaches 99.8% under normal walking(NM)conditions,99.1% under walking with bag(BG)conditions,and 88.2% under wearing coat or jacket(CL)conditions.Compared with GaitSet,it improves by 4.8%,11.9%,and17.8%,respectively;Experiments on the OU-MVLP dataset have achieved a Rank-1 accuracy of89.8%,which is 2.7% higher than the GaitSet.(2)Taking the keypoints of the skeleton as input,the gait information is decomposed into a subset of gait information for multiple parts,and an algorithm called PartGraph is proposed.Firstly,the algorithm uses a Keypoint Detection Network to obtain the skeleton keypoints corresponding to the gait silhouettes;Then,based on the fact that the gait patterns of different parts of the human body are different,the algorithm divides the skeleton data into blocks,divides them into corresponding keypoint sets for the parts,and inputs into the Parted Feature Extraction Network to extract features;Finally,these gait features are sent to the All-parts Feature Extraction Network for feature integration,and the final gait features are output.The algorithm is tested on a CASIA-B dataset.Its average Rank-1 accuracy in NM,BG,and CL conditions reaches 88.0%,76.8%,and 70.9%,respectively.The accuracy of the algorithm is improved compared to the Gait Graph under the same conditions..