Deep Learning Methods For Human Micro-Motion Features Recognition

In recent years,human body feature recognition has attracted extensive attention in the fields of security monitoring,automatic driving and remote health monitoring,in which the use of radar sensors can detect the micro-motion features effectively during movements.The micro-motion features of human body refer to the small movement information such as translation and rotation generated by the main body and its components.By extracting and classifying these micro-motion features,it can achieve the purpose of recognizing different human moving states.In this paper,we use the radar electromagnetic wave to obtain two types of micro-motion features in the process of movements and introduce the feature recognition methods based on deep learning.The recognition problems of different motions and human gaits under single-view condition and different motions under multi-view condition are studied respectively.The specific research works are organized as follows:Firstly,the micro-motion feature representation method of high resolution range change information is used to recognize different human motions.The two dimension time-range profiles of different human motions are obtained by detecting the range changes of torso and body in the process of movement.In the experiment,two kinds of human motions are collected and data sets are constructed.We utilize a convolutional neural network to extract features and recognize them.The results show that the proposed feature representation and recognition method for two types of 15 motions can achieve an accuracy rate over 90%.Secondly,the micro-motion feature representation method of gait micro-Doppler information is used for personnel recognition.Radars can detect the velocity changes of various body parts during walking,so that the corresponding micro-Doppler features can be obtained.Since the human body has memories for daily motions,the gait micro-motion can be considered as a unique identification tag for different people.In the experiment,we collect gait information of different people and adjust the structure of the convolutional neural network to train and test the datasets.The experimental results show that the recognition accuracy rate of gait features for 8 people is 96.9%,which also verify the feasibility of this method for personnel recognition.Thirdly,multi radar sensors are used to recognize two kinds of motions including walking and walking while holding an analog gun under multi-view conditions in a large area.Radars can only detect the radial scattering information under a single angle of view.When the motion path deviates from the radial direction of the radar,the lack of information will result in the misidentification of recognition results.We use wide-beam antennas and multi-point receivers to collect two kinds of motions in a large test area.We propose a feature fusion method which combines the convolutional neural network and traditional classification decisions to recognize the multi-point received micro-Doppler features.Experimental results show that the proposed fusion method can improve the stability of motions recognition in multi-view scenarios,and the recognition accuracy is 97.91%.