Multi-View Gait Recognition Based On Deep Generative Model

With the faster and faster development of social informatization,how to quickly and accurately identify a person’s identity has become a hot social issue.Today,when traditional identification methods based on text information are declining,the identification of each person’s identity information through biometrics has long been a research hotspot.The gait recognition technology aims to identify the identity of an individual through the walking posture.Because each individual has differences in bone length,muscle strength,joint flexibility,etc.,the gait has natural individual characteristics.Gait features are difficult to disguise and change,and the collection of features does not require the cooperation of subjects,and can be captured under long-distance conditions.These advantages make gait recognition have broad application prospects in all aspects of society.However,in the current training process of gait recognition model based on deep learning,due to the insufficient number of samples in training set,in the recognition task,the generalization ability of the model is weak,the true distribution of the data cannot be learned,and the multi-view recognition is not robust.The difference in gait characteristics of different perspectives of the same recognition object is often greater than the difference between the gait characteristics of different recognition objects under the same perspective.Aiming at the current problems of gait recognition that is not robust to changes in perspective,insufficient training sample set data,and the high cost of establishing large-scale labeled gait data integration,this paper proposes a method of using class conditional generative expressions.The method of expanding the sample set to increase the richness of the sample class has improved the accuracy of multi-view gait recognition.This article starts with different ways of expressing gait characteristics,and designs conditional generation models based on key point sequence and gait contour sequence.Specifically,it is difficult to make full use of the time sequence information during exercise and the problem that it is not robust to apparent changes in the method of using energy-like graphs as feature templates.In addition,this article uses a human body keypoint sequence that is more robust to changes in perspective and appearance.As input,a conditional generation model based on the human body’s key point sequence is proposed.Using category information as the condition,the multi-task joint training discriminant model and the generative model are used to constrain the generative model through the discriminant model so that it can accurately generate the keypoint sequence according to the category.Accurately and efficiently perform intra-class augmentation on each type of sample in the data set.Both the generative model and the discriminant model in multi-task training use LSTM as the skeleton network to model keypoint sequence.The generative model is based on the variational autoencoder of dynamic sequence,and discriminant model uses the triplet loss function,which can be fully utilized dynamic information improves the ability of discriminant model,and better guides the generation model to generate specified category data.Aiming at the problem that the key point sequence only models dynamic geometric information,does not contain apparent information,and has a small amount of comprehensive information,a class-condition generation model based on gait contour sequence is proposed to solve the problem of the vagueness of the data generated by the variational autoencoder.Based on the multi-task joint training of the generative model and the discriminant model,the generative adversarial training model is further integrated,and while accurately and efficiently augmenting each type of sample,it further improves the quality of the gait contour map sequence generated by the class condition.The proposed multi-task joint training model takes category information as the condition,so that the generative model can accurately control the category of the generated gait contour sequence.The modeled generative model can use discrete category conditions and sample continuous latent variables for each type of sample.Carry out enhancements to increase the diversity of changes within the sample category.Experiments prove that the method of augmenting the data set using the proposed generative model has good effects on a variety of gait recognition algorithms,and the recognition accuracy is improved.