| Gesture recognition is an important method in human-computer interaction.Traditional computer vision-based gesture recognition is not only susceptible to lighting effects,but also prone to problems such as user privacy leakage.Gesture recognition based on frequency mod-ulated continuous wave(FMCW)radar can solve the above problems,so it has a wide range of application scenarios.There are two main types of FMCW radars:(1)Radars with a large number of channels that can obtain point cloud data.(2)The number of channels is small,and the radar can only obtain raw data.Based on the above two radars,this thesis conducts gesture recognition research.The specific work is as follows:1.Research and propose a gesture recognition algorithm based on FMCW radar point cloud data.An effective gesture information is obtained by target detection,and the clustering algorithm combined with the pre-order memory queue is used to select the effective data frame.Feature fusion is used to solve the problem of inconsistent data points per frame,and feature compression is used to reduce the amount of data computation while retaining time dimension information.A lightweight neural network is designed for classification,and finally an accuracy rate of 98.21% is obtained on the collected data set,and the operation time of the algorithm is effectively reduced.2.Research and propose a gesture recognition algorithm based on FMCW radar raw data.Firstly,a pre-order root mean square error queue is used to select valid data frames.After filtering through distance and speed in two dimensions,a Moving Target Indicator(MTI)algo-rithm is used to filter out the static noise of Range-Doppler Image(RDI).Secondly,a method of adaptive fixed-time-length is used to solve the problem of inconsistency in the number of data frames,and data enhancement is performed on the small sample data set based on the character-istics of FMCW radar.the RDI and the Range-Angel Image(RAI)are fused as gesture features and classified through the designed neural network.In the end,an accuracy rate of 97.5% was obtained on the collected dataset,and an accuracy rate of 98.28% was obtained in Google’s open source deep-soli radar dataset,which was 11.11% higher than the algorithm proposed by the dataset.3.Research and solve related problems of gesture recognition in real-time reasoning sce-narios.Use multi-dimensional basic features to determine and solve the problem of false trig-gering.The problem of negative samples is solved by post-processing comparison of manually extracted features and outputting confidence.An accuracy rate of 90.5% was achieved in the real-time reasoning experiment,and finally the gesture recognition algorithm in this thesis was successfully applied to mobile terminal equipment. |
