Human Action Recognition Based On Microdoppler Signatures Of Millimeter-Wave Radar

Human action recognition technology has many applications in security and surveillance,medical and health care,human-computer interaction,and other fields.It is of great interest to academia and industry.Compared with traditional visual and wearable sensors,radar sensors have all-weather characteristics and high penetration characteristics and can protect the user’s privacy,so it is essential to study radar-based human action recognition technology.Based on millimeter-wave radar sensors,this paper focuses on human action recognition algorithms mainly from four aspects:human action simulation analysis,measured action data acquisition and processing,single-action recognition,and continuous action recognition.And the specific work is as follows:(1)Human action simulation analysis.Based on the human skeleton model,the influence of human action on radar echo signal is analyzed by simulating the whole process of executing a specific action in the corresponding simulation scenario.The micro-Doppler signatures are extracted using the time-frequency analysis method from the echo signal.By exploring the similarities and differences between the microDoppler signatures of nine types of daily human action,it was confirmed that the microDoppler signatures could be used to distinguish daily human action,which laid a theoretical foundation for the subsequent research on the measured movements.(2)Collection and processing of measured movement data.The human action data acquisition platform was built by combining four aspects: radar hardware selection,radar parameter setting,experimental scene arrangement,and real-world action description.Because of the energy dispersion phenomenon of the traditional timefrequency analysis method in processing the radar data,the RDM(Range Doppler Map)velocity dimensional projection method is adopted to construct the micro-Doppler time-frequency map frame by frame,and the feasibility analysis of this method is carried out.Finally,an upper computer software that can be used for real-time radar data acquisition and visual presentation of micro-Doppler signatures is designed and developed to provide actionable data for the subsequent research.(3)A single human action recognition method based on micro-Doppler signatures is proposed.Based on the radar measured data,11 micro-Doppler feature vectors with actual physical or abstract meanings were extracted from the micro-Doppler timefrequency map.Then,the support vector machine(SVM)after the adjustment and optimization of Bayes hyperparameters was used as the classifier to analyze the influence of each single feature vector and the combination of feature vectors on the classification accuracy.Finally,the best feature vector combination was obtained for human action recognition.The experimental results show that all the feature vectors extracted from the micro-Doppler time-frequency map can describe the micro-Doppler signatures intuitively,and the best combination of feature vectors was used to recognize the 9 kinds of actions of known individuals and unknown individuals,and the recognition accuracy is up to 96.06% and 91.44%,respectively.(4)A continuous human action recognition strategy based on micro-Doppler signatures is proposed.The strategy takes the improved Faster R-CNN model as the core,takes the micro-Doppler time-frequency map or video of continuous action as the input,recognizes the micro-Doppler signatures corresponding to every single action in the continuous action sequence with the powerful target detection ability of Faster RCNN network,and sets specific rules to output the prediction frame information,and processes the prediction frame information to obtain the start and end moments of each action in The starting and ending moments of each action in the sequence and its corresponding action category is obtained by processing the prediction frame information.The experimental results show that the Faster R-CNN model’s m AP(mean average precision)can be improved by 1.05% and 0.55% by modifying the backbone network and adjusting the anchor frame size in the RPN network,respectively.The continuous human action recognition strategy based on the improved Faster R-CNN can accurately obtain the action information under each time period in various continuous action sequences.Its average accuracy can reach 92.14%.Finally,the effects of feature-level noise and signal-level noise on detection results is further explored,and the excellent robustness of this strategy is verified.