Research On Detection And Classification Algorithm Of Gesture Target By FMCW Radar

Gesture is one of the most direct and vivid ways of human interaction.In recent years,gesture detection and recognition technology has gradually became a research hotspot in the field of human-computer interaction.The data sources of gesture recognition technology are mainly divided into wearable sensor technology,computer vision technology and wireless signal sensing technology.Human wearable sensors require users to wear special equipment,which is expensive and inconvenient to use.The camera based on visual technology may not work properly under the condition of strong light erosion and dark,and there is a risk of leaking user privacy.Frequency Modulated Continuous Wave radar has the advantages of convenience,strong robustness and high security.In our research,we proposed the research of gesture detection algorithm with FMCW radar.The specific research contents are as follows:Firstly,we discussed the detection of the segment of the gesture data and the method of parameter extraction.By analyzing the jitter characteristics of the gesture action on the amplitude of time domain,to cut the data segment of interest containing the gesture.Then,we improved the traditional parameter estimation method to extract the distance,doppler and angle spectrum,by extracting the gesture parameters.Secondly,we studied the noise and interference methods in doppler parameter estimation.Based on the range-doppler map obtained by radar signal processing,the noise removal is performed by the improved frame-difference method.Using principle of constant false alarm detection,the target detection is performed in the range-doppler map,and the interference target is suppressed by the characteristics of the gesture target.Thirdly,the multidimensional data construction method based on radar parameter extraction is studied.By coupling range,doppler and angle spectrum with time information,the range-time map,doppler-time map and angle-time map are constructed,and the timing information of gestures is mined by means of parameter-time map.A threechannel data set is constructed for the convolutional neural network.Finally,an end-to-end multi-branch convolutional neural network is studied.The multi-branch structure enables the model to simultaneously extract features of the rangetime map,doppler-time map,and angle-time map of the same gesture.Early fusion of the three features through the fully connected layer makes the network structure end-to-end(without step-by-step training).By collecting a large amount of gesture data,and processing the processed three image data into a multi-branch convolutional neural network for training and testing.The experimental results show that the classification accuracy of the six types of gestures reaches 95.33%,and the algorithm has the advantage of low complexity compared with the similar methods,which verifies the reliability and effectiveness of the proposed algorithm.