Research On FMCW Radar Human Activity Recognition Algorithm Based On Deep Learning

Human activity recognition(HAR)is a task aimed at identifying human actions using various general or specialized sensor data,typically using cameras and sensor data processing spatial or temporal data to understand human behavior.Compared to visual sensors,millimeter-wave radar has advantages such as all-weather,long-distance,and strong penetrability,which can achieve highprecision detection of Doppler,distance,and angle,as well as better privacy protection.Deep learning can automatically extract features end-to-end and output classification results compared to traditional hand-crafted feature-based algorithms.Based on this,this thesis proposes a behavior recognition algorithm based on deep learning and millimeter-wave radar to overcome the limitations of optical devices and traditional behavior recognition algorithms.However,existing millimeter-wave HAR algorithms mainly focus on the research of micro-Doppler feature maps and ignore the use of distance information contained in RF echo signals.Therefore,this thesis studies the feature fusion-based FMCW radar behavior recognition algorithm..The specific work of this thesis is as follows:(1)In response to the increasing complexity and number of classifications in radar-based behavior recognition,the performance of models that only use velocity information as input gradually decreases,leading to weakened differentiation effects between similar behaviors.This thesis addresses this issue by conducting research on data preprocessing,feature input,and network architecture,and proposes a multi-feature fusion behavior recognition algorithm.In terms of data processing,this study differs from traditional mean background subtraction methods by using a fourth-order Butterworth filter to remove DC and clutter,and using windowing to improve spectral leakage problems.For network feature input,this study inputs velocity and distance features into the network in a dual-stream manner,using four fusion methods including addition,multiplication,concatenation,and weighted scores.In experiments,this study compares the proposed model with existing models on publicly available datasets and self-built datasets in terms of accuracy,precision,recall,computational complexity,parameter volume,and inference time,thus verifying the performance and high recognition rate and generalization ability of the proposed model.(3)To address the problem of transferability in FMCW radar-based behavior recognition algorithms,which is affected by radar operating frequency bands,working scenarios,radar models,and different signal processing algorithms,this study proposes a cross-domain behavior recognition algorithm based on adversarial learning.In the first stage,this algorithm learns similar features between the source and target domains through weight sharing and maximum mean difference;In the second stage,the weights of stage one are used to initialize the feature extractors of the source and target domains,and the domain discriminator and target domain feature extractors are trained to compete in a game,allowing the target domain feature extractor to learn domain invariant features.This algorithm not only uses feature distance difference calculation to solve the problem of low-level feature differences,but also uses adversarial training to solve higher-level differences between features.In experiments,this study achieved cross domain human behavior recognition on both the self built radar dataset and the Glasgow public radar dataset,with no significant performance loss.At the same time,in order to verify the universality of the algorithm,this algorithm was compared with existing algorithms through cross domain experiments on three digital recognition image datasets:MNIST,USPS,SVHN,and Office31,verifying that the proposed algorithm model has higher recognition accuracy.