Gait Recognition Based On Deep Neural Networks

With the coming of the new artificial intelligence technology,deep learning-based neural network has been widely used in the field of biometric recognition.Traditional biometric features are usually collected using contact sensors or con-tactless sensors over a short distance,which are difficult to be employed to rec-ognize identity without cooperation in uncontrolled scenarios.Gait feature has many advantages,it's easy to be collected without cooperation,it's difficult to imitate or camouflage,it can be used with low image quality.Traditional gait recognition technology adopts handcrafted feature for gait recognition.The ac-curacy of gait recognition is very low due to view variations,wearing noise and carring things,which fails to meet the acquirement of actual application.There-fore,the deep learning-based gait recognition technology is employed to solve the problem of long-distance identity recognition.The research content of this paper are introduced as follows.1.An overview of gait recognition.This part consists of introduction and the overview of gait recognition technology.The former analyzes the generation background and research significance of gait recognition.Then it surveys the feasibility of gait recognition,it analyzes the development status and trend of gait recognition,the main problems exists in gait recognition.The latter analyzes the system architecture and hardware deployment of gait recognition,it introduces the system composition of gait recognition,the related methods and gait database.Finally,this paper employs metric learning,generation model and combined model to solve existing problems.2.Metric learning-based gait recognition.To make extra-class variations larg-er than inter-class variations,this paper designed three kinds of metric learning-based deep neural networks for gait recognition.CDNN takes gait sample pair as input,adopts softmax loss and verification loss to optimize network jointly.Triplet network adopts triplet sample set as input and takes triplet loss to optimize network.Besides,we reconstruct triplet net-work by adding LSTM unit after convolution layers.Experimental results show that these methods can improve gait recognition precision effectively.3.Combined model-based gait recognition.To effectively apply the tempo-ral feature,a spatial-temporal deep neural network(STDNN),consists of spatial feature network(SFN)and temporal feature network(TFN),is pro-posed in this paper.SFN is adopted to extract the static feature of gait samples used for gait recognition.TFN is adopted to extract the temporal feature for gait recognition.Finally,a special fusion strategy is adopt to fuse the recognition scores achieved by SFN and TFN.The best recognition score reaches 95.67%.Compare with the existing method,it rises 4.25%.Experimental results turn out that STDNN is capable of overcoming the influence of wearing noise and view variations on gait recognition accuracy.4.View transformation model-based gait recognition.To overcome the influ-ence of viewpoint variations,this paper adopts stacked sparse autoencoder(SSAE)network to transfer the view domain feature of the testing sam-ples,and generates the virtual sample with the same view domain feature as the training samples.Adopting the virtual sample for gait recognition is capable of overcoming the influence of view variations and improving the accuracy of multi-view gait recognition.Its best recognition score reach-es 93.67%,which outperforms the current best value greatly.Extensive experimental results indicate that SSAE outperforms traditional method.Finally,we summarized full paper and predicted the research trend in future.This paper analyzes its comtributions and innovations,it analyzes the existing problems and points out the research direction in the future.