Research On Indoor Human Action Recognition Algorithm Based On Multi-Antenna FMCW Radar

With the popularity of the intelligent home,human action recognition becomes a research hotspot.The use of radar sensors as an instrument to detect human targets will not burden the user,nor will it be restricted by environmental illumination conditions.At the same time,the high accuracy of radar measurement provides a good foundation for human action recognition.In addition,with the development and popularization of machine learning,the method and the accuracy of human action recognition has been improved.Using radar sensors to classify and recognize dynamic human behaviors,the general steps are to collect radar data,process radar signals,extract feature information,and classify actions based on features.The analysis shows that dynamic human behavior firstly includes three-dimensional spatial characteristics,which express the position and shape of the human body target,and secondly include time characteristics,which express the change characteristics of the target position and shape over time.The existing implementation schemes of action recognition have this following problems:firstly,the radar signal processing algorithm has high computational complexity;secondly,the noise and interference signals in the parameter spectrogram may affect the feature information extraction;thirdly,insufficient feature extraction may lead to reduced recognition accuracy.Based on this,a multi-antenna Frequency Modulated Continuous Wave to realize the classification and recognition of dynamic human behavior is used in this thesis.The main research contents of this thesis are as follows:Aiming at the first problem,the Fast Fourier Transform(FFT)algorithm and the Minimum Power Distortionless Response(MPDR)beamforming algorithm are introduced to instead Multiple Signal Classification(MUSIC)algorithm to estimate distance and angle parameters.Then uses FFT-MPDR algorithm to achieve joint estimation of two-dimensional parameters.Finally compares and analyzes the performance of FFT-MPDR,FFT-MUSIC,2D-MUSIC algorithm.Aiming at the second problem,for parameter spectrogram generated by the FFT-MPDR algorithm,the noise and interference signal elimination algorithm based on multi-sweep frequency signals is adopted.This algorithm combines multiple frequency sweep signals to eliminate the environmental noise and the interference signal reflected by irrelevant objects in parameter spectrogram.Aiming at the third problem,a multi-dimensional feature fusion neural network is designed and constructed in this thesis.This network uses two independent Convolutional Neural Networks to extracts horizontal and vertical spatial features of human action separately for fusion,and then uses one recurrent neural network to extract the temporal features of the fused spatial features,and realizes classification through the temporal features.The network is trained and verified through the human behavior dataset to realize the recognition of 8 human behaviors with an accuracy rate of 97.29%.