Research On Gait Recognition Algorithm Based On Feature-Enhanced GaitSet And Micromotion Features

Through the principles of computer and biostatistics,the technology of using the unique physiological characteristics and behavioral features of human body for personal identification is called biometrics.Biometrics is becoming an important scientific field as the security issues in people's daily life with the increasing security concerns in people's daily lives.Gait recognition can monitor people from a distance and become one of the key research directions in the field of computer vision and pattern recognition.It's difficult to imitate and expected to be applied in criminal investigation,access control and other scenarios.Computer vision has grown by leaps and bounds with deep learning in recent years.Gait recognition methods based on deep learning also have made rapid progress.Firstly,the paper starts from the background and significance of gait recognition research and elaborates the development origin and realistic needs of gait recognition technology.Also,paper explains that gait recognition technology has the characteristics of non-contact and concealment compared with other biometric technologies,and analyzes the current research status of gait recognition technology in the research field.The difficulties that still exist in gait recognition technology.On this basis,we introduce the convolutional neural network,batch normalization,activation function and back propagation algorithms according to the latest deep learning methods widely used in gait recognition technology,and describe the two mainstream technical methods of gait recognition.Then,the paper proposes an Improved Feature Enhancement Gait Set(IFE-Gait Set)gait recognition algorithm.The algorithm uses the Open CV framework and morphological processing to crop and align,normalize and repair the CASIA-B gait contour dataset to obtain a better processed-CASIA-B pedestrian gait contour dataset,which still has more voids and misconnected areas than the original gait video dataset;using three-channel convolutional feature fusion to enhance the feature extraction capability of Gait Set network,and using the improved inter-class optimized triplet loss function for penalty training in the training phase to close the distance of the same label in the feature space and furtherly enhancing the network learning capability.Through experimental validation of the optimized network model,the IFE-Gait Set was trained in the Processed-CASIA-B pedestrian gait profile dataset using inter-class optimization triplet loss,and the average first hit recognition accuracy was improved by 1.91%,3.281%,and 2.595% in the normal state,coat wearing state,and backpack carrying state.Respectively,the overall average first hit(Rank-1)recognition accuracy for the three states is improved by 2.6%,and parameters of the network model are only increased by 6%.At the same time,the algorithm performs ablation experiments under data set processing and loss function optimization,and the experimental results show that a more obvious accuracy improvement is obtained.Finally,the paper proposes a Gait Recognition on Micromotion Features(GRMF)gait recognition algorithm based on micro-motion features.The algorithm borrows the idea of local features in pedestrian re-recognition and focuses on the information of tiny motion amplitude in pedestrian gait to propose micro-motion features.The algorithm first performs the cropping alignment,normalization and patching operations on the original data set,and then implements the micro-motion feature extraction and feature mapping based on convolutional neural network.Among them,the micro-motion feature extraction is done by the cut block convolution with the help of residual structure on the gait profile set,and then the micro-motion features are sent to the feature mapping module to map to the pedestrian gait feature vector through the set pooling layer,and finally the penalty training is performed by using the inter-class optimization triad loss,and the final gait recognition accuracy results are obtained by testing in the test set.The GRMF gait recognition algorithm achieves an average first hit(Rank-1)recognition accuracy of 95.336%,90.233%,and 75.782% in the normal state,coat state,and backpack state on the Processed-CASIA-B dataset,especially in the coat state(when the pedestrian gait features are substantially obscured by the coat),demonstrating the effectiveness of micromotion features.