Research On Single-View And Multi-View Gait Recognition Algorithms

Gait recognition is a technique that uses gait information to identify a person's identity.On the one hand,the learning performance of gait recognition algorithms based on deep learning relies seriously on careful tuning of parameters,and training a deep learning model requires a large amount of training data,while the existing gait database is limited in scale.Therefore,exploring gait recognition algorithms suitable for small data sets with few parameters has great scientific value and practical significance.On the other hand,the gait recognition algorithm based on convolutional neural network has achieved excellent recognition performance on multi-view gait recognition,and feature extraction is crucial for gait recognition,but the extraction of gait features is affected by walking angle,carrying and wearing conditions,so it is of great significance to use convolutional neural network to extract abundant and representative multi-view gait features.In view of the above problems,this thesis proposes a single-view gait recognition algorithm based on Siamese Deep Forest(SDF).This method requires fewer parameters and less sample size,which provides another way to solve single-view gait recognition problem.In addition,this thesis innovatively applies the highway network to the multi-view gait recognition field,effectively improving the accuracy of multi-view gait recognition.The main research contents and results of this thesis are as follows:(1)Aiming at the problems that gait recognition algorithm based on deep learning are with too many parameters and requires a large amount of training data,a single-view gait recognition algorithm based on SDF is proposed.Firstly,by analyzing the training method of SDF,combined with gait recognition problem,a single-view gait recognition dataset suitable for SDF is constructed;Secondly,by discussing the weight calculation method,loss function structure and training algorithm of SDF,the training and test process of single-view gait recognition algorithm based on SDF are designed.In order to enable SDF to separate a pair of gait energy image(GEI)with sufficient confidence and further improve the generalization ability of SDF,we propose a modified single-view gait recognition algorithm for Siamese Deep Forest(MSDF)by introducing classification intervals.The experimental results show that the SDF and MSDF structures can be used to solve the single-view gait recognition problem,and the algorithm requires less parameters and less sample size.(2)Aiming at the problem that multi-view gait recognition needs to extract rich features,two kinds of multi-view gait recognition methods based on highway network are proposed.Firstly,by rethinking the design motivation of the highway network,we considered that highway network consists of two feature extraction modules.Secondly,we design two types of multi-view gait recognition network structures based on highway network are according to the previous assumptions,discuss the similarities and differences of feature extraction methods of these two types of networks in the mean while.Subsequently,we construct a multi-view gait dataset and the accuracy of the multi-view gait recognition of the two types of networks is respectively tested on the dataset.Finally,the optimal model is selected for multiple sets of experiments.The experiment proves that the highway network structure designed in this paper can extract rich gait features and can improve the accuracy of multiview gait recognition.