Research On FMCW Gesture Recognition Algorithm Based On Time Series Features

Human Computer Interaction(HCI)means that users can interact with devices using simple gestures to let computers understand human behavior.As an important way in the field of human-computer interaction,gesture recognition has gradually become a hot research direction in this field.As a tool for detection and positioning,radar has the advantages of high precision and low interference compared to other sensors.Gesture recognition generally collects raw data,then extracts the characteristics of each type of gesture,and finally classifies the model by feature.The existing gesture recognition mainly has the following characteristics: First,after calculating the radial motion parameters of the gesture relative to the radar,there is no further attention to the lateral change of the gesture,so the parameter information of the representation gesture is insufficient;Second,in addition to the features extracted at each instant,there are time dimensions since the gesture is an ongoing process,which makes the parameter dimension larger and the feature extraction difficult.Third,each type of gesture has unique characteristics according to its chronological order.However,after each instantaneous feature extraction,the timing feature information has not received the attention it deserves.In summary,the gesture recognition algorithm based on Frequency Modulated Continuous Wave(FMCW)radar signal is proposed in this paper.The main work of this thesis is as follows.Firstly,the intermediate frequency(IF)signal of the FMCW radar is used to calculate the range,speed and angle parameters of the gesture.A multi-frame Range-Doppler Map(RDM)is obtained according to a 2D Fast Fourier Transform(2D-FFT)algorithm,and then multiple signal classification(MUSIC)algorithm is used.Angle was calculated by this algorithm and a single-frame angle time map(ATM)is generated through multi-frame accumulation.Secondly,according to the characteristics of the gesture peaks in RDM and ATM,the adaptive gradient of each column of pixel values in the graph is calculated,and the peak interference filtering in RDM is completed by first-order exponential smoothing,and then RDM and ATM are performed by wavelet transform.Decomposition and reconstruction implement Image Enhancement(IE).Thirdly,we build a deep learning neural network for gesture feature extraction and classification.According to the characteristics of the gesture parameter map,the Inflated 3D Network(I3D)and the Convolutional Neural Networks(CNN)are used to extract the spatiotemporal features,and then the feature results are recombined,and then the long short-term memory(LSTM)networks are used to extract the range,speed and angle change information of the gesture.