Research On Through-wall Radar Human Motion Recognition Method In Occluded Scene

The through-wall radar obtains the motion information of the occluded human target by emitting electromagnetic waves that can penetrate non-metallic walls,and uses signal processing and feature transformation to extract and analyze the feature differences of different motions,realize the recognition of human motion,and can always perceive the state of the occluded space target,providing technical support for medical monitoring,disaster rescue and tactical decision-making.However,the resolution of the through-wall radar is limited,and the characteristics of human motion acquired are prone to aliasing.At the same time,due to the complex influence of the wall and the temporal and spatial variability of human motion,how to effectively realize the robust and real-time recognition of human motion by the through-wall radar is an important and challenging problem.This thesis focuses on the above-mentioned problems of real-time recognition of human motion by through-wall radar,and carries out research work on through-wall radar echo modeling and preprocessing of human motion,robust real-time recognition method of human motion,and multi-task real-time recognition method of through-wall radar and verified with the measured data.The main work is as follows:1.The radar human motion echo modeling,signal preprocessing and feature extraction methods are studied.Based on the visual motion capture data,the human body and the human motion echo of the through-wall radar is simulated,and a reasonable data processing method is designed to extract the multi-domain features of human motion through time-frequency analysis and frequency fusion,realizing effective characterization of wall shielding human motional features.2.The time-frequency analysis method of human motion based on CEEMDAN and multi-frequency fusion is studied.By mode decomposition and spectrum fusion of multiple single-frequency data,the limitation of time-frequency resolution of traditional STFT analysis method is effectively avoided,and the high-resolution time-frequency feature of human motion are capable to be extracted.3.The method of human motion sequential recognition based on convolutional recurrent neural network is studied.The convolutional GRU is used to extract the temporal-spatial features of human motion distance-doppler-time three-dimensional data,and the robust sequential recognition of wall shielding human motions by through-wall radar is effectively realized.4.The multi-feature robust real-time recognition method and multi-view robust realtime recognition method based on ensemble learning are studied.Different sequential recognition models are constructed by using multiple features,and frame-level recognition results are integrated in the time dimension,which effectively solves the problem exists in single-view robust real-time recognition.Secondly,the GRU recurrent neural network is used to extract and fuse the human motion features of multi-view through-wall radars,and the feature level integration recognition is realized through random forest,implementing the robust real-time recognition of multi-view human motion.5.A robust real-time recognition method based on multi-task learning is studied.Through the effective combination of motion recognition and identity recognition tasks,it is solved that a single model can realize two real-time recognition functions at the same time,which realizes the robust recognition of human motion and target identity,and improves the recognition performance of the model.