| Biometric identification has become one of the main methods in the field of personal identification due to its high accuracy,convenience and speed.Gait recognition is also a biometric identification method,which recognizes the identity by analyzing the body structure and walking posture of the target person.Due to its unique advantages,it has gradually developed and has been brilliant in the field of criminal investigation.A complete gait recognition system includes three steps:gait information acquisition,gait feature extraction and gait feature matching.The computer vision technology based on deep learning is very compatible with the gait recognition task,which promotes the progress of gait recognition technology.However,the practical application of gait recognition is still limited by many problems,such as view diversity,carrying objects or clothing changes,occlusion,and unstable walking speed.In view of these problems,this paper conducts an in-depth study on gait recognition algorithms:Many studies have shown that partitioning the gait feature map can improve the accuracy of gait recognition.However,most models just cut the feature map at a fixed scale,which loses the dependence between each body-parts.So,this paper proposes a structure called Part Feature Relationship Extractor(PFRE)to discover all of the relationships between each part for gait recognition.PFRE combined with Convolutional Neural Network(CNN)can form a gait recognition model RPNet.PFRE is divided into two parts:the Total-Partial Feature Extractor(TPFE)is used to extract the features of different scale blocks,and the Adjacent Feature Relation Extractor(AFRE)is used to find the relationships between each block.RPNet is tested on three public gait datasets,CASIA-B,OU-LP and OU-MVLP.And it exhibits a significant level of robustness to occlusion situations.Based on RPNet,this paper designs a more advanced gait recognition model GGCN,which uses multi-type gait sequences as input and eliminates the effects of various covariates through a supervised mapping module.The GGCN processes the gait sequence in three steps.First,the supervised mapping generate network is used to extract low-level features and remove the features generated by interference.Then,the low-level features are put into the encoder network to obtain high-level features.Finally,the high-level features are put into the feature mapping network to acquire more recognizable features.Experimental results on public gait datasets show that GGCN effectively improves the recognition accuracy of RPNet.The binary sequential inputs determine that the key-point of gait recognition task is how to design a framework to extract multi-scale spatiotemporal information and form efficient gait features.However,previous studies mainly focused on coupling information flow into more complex and deeper networks,lacking in-depth analysis of the synergistic mechanism of different signals.This paper proposes a unified feature learning network named GaitU.GaitU is a fully decoupled gait recognition network in spatial-temporal and scale,which can effectively extract coarse-grained and fine-grained features in both spatial and temporal domains.GaitU significantly advancing the SOTA results for gait recognition.The application scenarios of gait recognition include outdoor natural environment.However,there are few researches on gait recognition in open environment.This paper relies on the two latest outdoor gait datasets GREW and Gait3D to explore gait recognition in open environments under the GaitU framework.This paper proposes three gait recognition algorithms,solves some difficulties in gait recognition,and explores gait recognition algorithms in open environments,which can provide some help for the construction of gait recognition systems. |
PDF Full Text Request