Research On Hand Gesture Recognition Algorithm Based On Millimeter-Wave Radar

Being as an information media with the ability of delivering information directly,hand gesture has broad application in human daily life,and in the field of HumanComputer Interaction(HCI).Due to that,Hand gesture recognition technology is getting increasing attention these years.The millimeter-wave-radar-based hand gesture recognition technology has plenty of advantages like having no restriction of light condition,high privacy and easy deployment,and can be adapted easily to Advanced Driving Assistance System,Internet of Things,smart home,etc.Therefore,it is of great significance to study the radar-based hand gesture recognition technology.Due to the near-field distance of usage of hand gestures,and the hand belonging to the flexible object,hand gesture recognition technology using radar has exceptionalities compared with the traditional fields of radar usage.The critical problems of the technology are how to extract features of the radar data reflecting from single hand gesture efficiently and accurately,and designing proper classification algorithm adapted to the typical hand gesture recognition problem.Focusing on these problems,this thesis carried out the research of hand gesture recognition algorithms based on frequency modulated continuous wave(FMCW)radar.The main contents are as follows:1.The signal preprocessing procedures including clutter suppression are studied.The database of eight time-varying hand gestures,consisting of 3200 sample data is built,by using 77 G millimeter-wave radar with 1TX and 4RX,which provides the foundation for the researches in the subsequent sections.2.The multi-dimensional feature extraction method is studied.By using pulse compression,short-time Fourier transform and the direction of arrival estimation,the range information,doppler information and azimuth information of the hand gesture are obtained.3.Aiming at the problem of feature fusion,the CNN based hand gesture recognition algorithm combining with attention mechanism is proposed.The introduction of selfattention block add context features to the origin,contributing to the feature fusion and the raise of the recognition accuracy.4.Aiming at the problem of inadequate usage of data and the high extracting complexity using man-made features,as well as the problem of heavy computing burden of the traditional CNN method,the LSTM based end-to-end hand gesture recognition algorithm is proposed,in which hand gesture data is processed in one-dimensional form.The feature extraction part is all contained in the algorithm.With the high performance of recognition,the algorithm has marked low complexity.The methods above are verified by the theoretical analysis and real data experiment.The results shows that the methods and algorithms proposed can achieve adequate and efficient feature extraction,and the accurate hand gesture recognition.