Research On The Recognition Of Hidden Human Posture Using SFCW Radar Based On Deep Learning

In recent years,the non-contact wall penetrating detection and classification technology of human postures has significant application values in many fields such as public security and protection,counter-terrorism actions,disaster assistance and so on.Most of the researches of hidden human postures recognition investigate the class of human postures in a fixed time window simply based on a single aspect of the range map or Micro-Doppler features.In this way,there are some problems such as insufficient feature information extraction and difficulty in distinguishing incomplete action or two kinds of actions in a fixed time.Therefore,this paper adopts the method based on deep learning to carry out research on the above problems,as follows:Aiming at the problem of the application principle of non-real-time recognition of the whole description model of the human postures of the wall-penetrating radar,a method of modeling the hidden human postures based on SFCW radar is proposed.Firstly,a discrete human postures simulation model is constructed to realize the feature description of the hidden human postures,so as to provide a model basis for the non-real-time recognition of the hidden human postures.Then,the simulation radar echo of the human body is constructed based on the SFCW radar system,which provides the data basis for the feature extraction of human postures.Finally,IDFT and clutter suppression were used to generate slow time range map,and STFT was used to generate slow time Micro-Doppler map,which provided the data basis for the subsequent deep learning based research on hidden human posture recognition,which provide the data basis for the following researches on hidden human postures recognition based on deep learning.The results of simulation and actual measurement show that the method can extract the range map and the Micro-Doppler features of human posture in the scene.Aiming at the problem that it is difficult to reflect the position information of human joints by only using range map in the research of hidden human posture recognition,and that only extracting Micro-Doppler features sometimes overlays the features with not obvious radial velocity,a multi-dimensional parameter human posture recognition method based on IBRes Net is proposed.Firstly,a feature form of three-dimensional tensor is established to expand the feature dimension of hidden human postures,to effectively utilize the range map features and Micro-Doppler features of human postures.Then a neural network based on improved bottleneck residual module is designed to realize the integration of feature extraction and feature recognition.Finally,the training parameters of the neural network are adjusted to improve the recognition efficiency and accuracy.The experimental results show that the accuracy of this method is 4 to 7 percent higher than that of range map or Micro-Doppler map alone.Aiming at the problem that the research of hidden human posture recognition is based on the whole human body behavior,which needs to extract the feature spectrum of fixed time period for recognition,a real-time recognition method of multi-dimensional parameter human posture based on 2C-LSTM network is proposed.Firstly,the slow-time range map and slow-time Micro-Doppler map of human body posture are constructed,the type labels of each slow time are marked and empty labels are added to represent the static state.Then,the multi-dimensional parameters are reduced and fused by the full connection neural network of two channels.Finally,the real-time recognition of human body posture is completed by two layers of the long-short memory cells?one full connection layer and Softmax classifier.The experimental results show that the network has the ability of short-term decision of human posture categories.After the first 15% information of perception behavior,it can accurately identify the posture categories,and the accuracy can reach 93.38%.