| Gait,as a complex spatiotemporal biometric feature,is widely applied to human recognition due to its difficulty in camouflage and feasibility to be captured at a distance.Current research on gait recognition is mostly based on RGB cameras.However,this intrusive method based on Computer Vision often collect users’ private data,which raises privacy and security concerns.Millimeter-Wave-based Radio Frequency method can protect users’ privacy comparatively,and has strong directionality and high resolution,which makes it an ideal substitute for gait recognition.In this thesis,millimeter-Wave(mm Wave)radar is used to collect raw signals,and on this basis,gait features are extracted to realize human recognition.The Range-Doppler features are extracted from the data collected by the mm Wave radar frame by frame,and background denoising is performed subsequently to construct an environment-independent spatiotemporal gait dataset.Multiple models based on Deep Learning are proposed to analyze the spatiotemporal sequences,and DS evidence theory is employed to fuse the results of multiple classifiers to improve recognition accuracy.Main work and innovations of this thesis are as follows:(1)The users’ gait is collected by mm Wave radar,followed up by signal processing to obtain the corresponding Range-Doppler features.Based on this,an environment-independent spatiotemporal gait dataset containing 11,600 samples is constructed.(2)Two different spatiotemporal network models,Att Res Net-LSTM and Conv RNN-Res Net,are put forward to resolve the spatiotemporal sequence.Att Res Net-LSTM adopts a Residual Network(Res Net)integrated with modules with attention mechanism to extract the spatial features.A one-dimensional intermediate code for each timestep is generated,which is thereafter sent into a Long Short-term Memory(LSTM)network for classification.Conv RNN-Res Net exploits the Convolutional Recurrent Neural Network(Conv RNN)to extract spatiotemporal features to generate an intermediate spatiotemporal code,which is subsequently fed into the Res Net for classification.(3)Based on DS evidence theory,a modified Yager’s rule is adopted to fuse the prediction of different classifiers.The accuracy is increased from 94.5% to 95.17% on the test set under a 2.5s observation period which has a poor performance under a single classifier. |
