Study On Micromotion Feature Extraction And Recognition Techniques For Human Gaits

When moving along the radar line of sight(LOS),the target or some structures on it are vibrating or rotating in addition to the bulk motion,known as the micromotion dynamics.Micromotion will induce additional frequency modulation around the main Doppler spectrum corresponding to the bulk motion on the returned signal,and generate microDoppler(m-D).In addition to the bulk motion induced by the torso,some human structures,e.g.,the limbs,are rotating or swinging during observation,which could induce m-D signatures on radar echoes.Because different m-D signatures usually correspond to particular human gaits,they serve as effective features for radar-based human gait recognition.Due to the characteristics of all-day,strong environmental adaptability,longaction distance,and non-contact,study on m-D feature extraction and recognition of human gaits is of great importance in military and civilian fields.Much progress has been made in m-D feature extraction and recognition of human gaits,but there still exists some challenging problems,i.e.,complex micromotion modeling,feature extraction in complex observation environment,and each dimensional feature utilization.With the support of the National Natural Science Foundation of China,the Joint Fund of the Ministry of Education for Equipment Preliminary Research,and the 13 th Five-Year Plan Pre-Research and other projects,this dissertation addresses the above-mentioned difficulties and key issues,and carries out in-depth research work on feature extraction and recognition of micromotion targets.The main content of this dissertation can be summarized into five parts as follows.1.For non-rigid human with complex structure and variable micromotion forms,a radar echo generation method based on motion capture data of human gait is proposed to address the difficulties in parametric modeling and in obtaining a large amount of public real data of radar,etc.Then the dataset of human gait is constructed,which lays the foundation for subsequent research on complex micromotion recognition for human gaits.2.To cope with the problem of severely defocused with spurious peaks in the traditional JTF analysis method under the complex observation environment,two sparse timefrequency reconstruction methods for micromotion targets are proposed.First,based on the linearity of Short-Time Fourier Transform(STFT),a sparse observation model in the joint time-frequency(JTF)domain of the micromotion target is constructed.Then a JTF reconstruction method is proposed by combining variable splitting with alternating direction method of multipliers(ADMM),which can obtain high-quality JTF distributions under data corruption scenarios.To address the problem of large reconstruction error under the low signal-to-noise ratio(SNR)scenarios,a probabilistic model is constructed by utilizing the priori information of the target and environment,and then a JTF reconstruction method based on single-window variational inference(SW-VI)is proposed.Compared with traditional JTF analysis methods such as STFT,the proposed methods can obtain better-focused JTF images of micromotion targets in complex observation environments.3.To address the problem that existing deep neural network based methods only use single window-length JTF images for recognition,which leads to insufficient extraction of micromotion features and low recognition accuracy,a human gait recognition method based on Dual-Channel Deep Convolutional Neural Network(DC-DCNN)is proposed.First,the JTF images with different STFT window lengths are input into the dual channels,and the torso features and limb features are extracted independently.Then the features of the two channels are spliced and fused to utilize the features extracted from the JTF images with different window lengths simultaneously.Finally,output the recognition results.The recognition results of human gait data show that DC-DCNN improves the recognition accuracy by 3% without increasing any radar resources compared to the single-channel network with only single window length.4.In this part,the multi-spectral channel attention based dual-channel network(MSCADCN)is proposed to avoid the degradation of recognition performance due to redundant features extracted by deep convolutional neural networks.The MSCA-DCN achieves channel domain enhancement and redundant information suppression of feature maps by weighting each channel using the multi-spectral channel attention module.Then the hybrid attention based dual-channel network(HA-DCN)is further proposed to enhance the important information in both spatial and channel domains of feature maps and suppress the multi-dimensional redundant information by combining spatial attention and multispectral channel attention.The experimental results show that the recognition accuracy of human gait data is improved by 2.4% and 2.8% for MSCA-DCN and HA-DCN,respectively,compared with the DC-DCNN method.5.To address the shortcomings of the existing methods of micromotion recognition for human gaits,which only use single-modal radar data such as JTF images and do not fully exploit time-domain features of the target,the homo-modal and cross-modal feature fusion network(HMCM-FFN)is designed.HMCM-FFN simultaneously exploits both JTF domain and time-domain information of micromotion target echoes.In particular,the network achieves information interaction between the same modal data and information fusion between different modal data through homo-modal and cross-modal feature fusion modules,respectively.The recognition results of the three types of human gait data show that HMCMFFN improves the recognition rate by 4% compared with the DC-DCNN method without increasing the radar resources,and has robustness to noise.