Human Motion State Classification Algorithm Based On Linear Frequency Modulated Continuous Wave Radar

LFMCW radar system has simple structure and good close detection performance.Compared with other sensors,the human detection based on LFMCW radar has advantages of all-weather and strong anti-jamming ability.Using LFMCW radar to achieve highprecision classification of different human motion state can detect people with abnormal behaviors in public places and ensure the safety of citizens.It can also be used for on-board products to keep a safe distance between vehicles and pedestrians.The classification of human motion state based on LFMCW radar has a wide application prospect in antiterrorism,traffic safety and human motion study.In this paper,Infineon 24 GHz continuous wave radar is used to carry out multiple group walking and running measurements for people of different heights and genders,and collect a large amount of data.After preprocessing the radar original data,time-frequency analysis method is used to obtain the time-frequency graph of human walking and running.Machine learning and deep learning methods were used to classify the time-frequency graphs of human walking and running to improve the classification performance of the model.First,the time-frequency graph of walking and running is classified by machine learning method,The HOG feature,LBP feature and HAAR feature of time-frequency graph were extracted.The SVM classifier was selected and a single feature was used to train SVM respectively to obtain three SVM models.The performance differences of SVM models with different feature training were compared and analyzed.In order to make full use of the timefrequency diagram information,the model is trained by different feature combinations to improve the model performance.Then XGBOOST classifier is selected,tree depth is adjusted,and the XGBOOST model with the best performance is obtained by comparison experiment.Finally,the method of stacking is proposed to fuse SVM model and XGBOOST model,The optimal fusion mode was determined through comparative experiments,and the performance of the model was tested by the measured data.Compared with the model of single feature training and the model of combined feature training,the model performance was improved.When the time-frequency graph of walking and running of human body is classified by machine learning method,because of the diversity of feature extraction and the diversity of classifier selection,the process is relatively complex and the final model performance is not optimal.Therefore,the time-frequency graph is classified by the method of deep learning.According to the simple structure of time-frequency graph,classification by complex networks may lead to overfitting,so the shallow convolutional neural network is designed first and the network is optimized from both data and model to improve the performance of the model.Then the network structure and layer number are changed and the optimal network structure and layer number are determined through comparative experiments.In this paper,different methods are adopted to classify the time-frequency graphs of human walking and running,and the performance of the model is gradually improved through optimization,design and improvement.Finally,the classification accuracy of the model on test set can reach 97.17%.