Identity Authentication And Recognition Technology Based On Inertial Sensor

With the rapid development of Internet of Things technology,wearable smart devices(WID)such as smart phones,smart watches/bands,smart glasses,smart clothing,etc.,have been widely used in our daily life,and play important roles in the fields of entertainment and social contacts,instant messaging,mobile payment,navigation and positioning,health monitoring,etc..As WID collect and store large amounts of personal sensitive information during use,their security have attracted more and more attention.At present,biometrics are the latest techniques for access control of WID.Gait recognition,as one kind of emerging biometrics,can recognize or authenticate people's identity through the unique posture they walk.Compared with traditional biometrics such as face or fingerprint recognition,gait recognition technology based on the built-in inertial sensor of WID has the advantages of active,real-time and continuous recognition.It can be used as an effective method to secure the information stored in WID and has great research and potential application value.Gait recognition technology for access control of WID has gradually become a research focus of scholars at home and abroad.Existing gait recognition methods,including template matching based methods and machine learning based methods,cannot yet meet practical application requirements in terms of robustness and accuracy.The accuracy of recognition and authentication are dramatically reduced in complex and dynamic wild world scenarios.With the improvement of the computing power of WID,the use of deep learning networks with powerful representation learning capabilities for gait recognition research has gradually become a new trend.To improve the robustness as well as reduce the model complexity of gait recognition based WID access control algorithms in real scenarios,this thesis designed two different deep learning models for multi-user systems to realize the gait-based identification and authentication.The main work and contributions of this paper listed as follows:(1)An identity recognition method based on lightweight attention network was proposed,which was used to identify the specific identity of legitimate users.In this work,a four-layer lightweight convolutional neural network was first designed to extract gait features.Then,a new attention weight calculation method based on contextual encoding information was proposed.Based on the proposed weight calculation method and depthwise separable convolution,this thesis designed an attention module which was integrated into the lightweight convolutional neural network to enhance gait features and simplify the complexity of the model.Finally,the Softmax classifier was employed for classification based on the enhanced gait features.This thesis conducted a series of comprehensive experiments to evaluate the performance of the proposed method on the gait dataset collected in real scenarios.The effect of the attention mechanisms,different data segmentation methods,and different attention mechanisms on gait recognition performance was studied and analyzed.The comparison results with the existing similar research in terms of recognition accuracy and model parameters shown that our proposed lightweight attention network achieved higher recognition performance when the model parameters were reduced by 87.8% on average.(2)An identity authentication method based on residual Auto Encoder network was proposed,which was used to verify the legitimacy of the user's identity and prevent unauthorized users from intruding into the system.This work was performed on the basis of the first work.Since the training data used only contains the gait data of legal users,this thesis first used the proposed lightweight attention network to extract gait features with individual discrimination from the training data.Then,the extracted features were input to the residual autoencoder network for unsupervised training.Finally,the trained residual Auto Encoder network was used to reconstruct the new sample need to be tested.By comparing the loss value between the original sample and the reconstructed sample to be tested,it can be determined whether the identity of the new sample was legitimate.In the experimental part,this thesis used the gait data set collected in real scenes to conduct experiments and compared with the methods proposed by existing researches.The experimental results shown that the method proposed in this thesis has achieved better certification performance,with an average reduction in EER by 4.36% and an average increase in AUC by 3.59%.