Research On FMCW Radar Gesture Recognition Algorithm Based On Deep Learning

Gesture recognition is a natural and intuitive way of human-computer interaction,and users can control electronic devices through hand movements.As the key technology of 5G wireless communication,millimeter wave can greatly improve the speed of wireless network.In addition to ultra-high-speed wireless transmission,the short wavelength,large bandwidth,and directional beam characteristics of millimeter waves also make it possible to recognize human gestures with high resolution and high robustness.In order to achieve a more intelligent and convenient human-computer interaction experience,this paper conducts research on gesture classification and recognition based on the FMCW millimeter-wave radar platform.The specific work is as follows:(1)Aiming at the problem of poor classification of tiny finger movements and confusing gestures,this paper studies gesture data preprocessing,feature utilization and network architecture,and proposes a gesture recognition algorithm based on multi-scale feature fusion.In terms of data preprocessing,the echo signal is preprocessed to eliminate abnormal points and invalid points,so as to achieve the effect of suppressing clutter and noise.In addition,using the characteristics of multiple transmissions and multiple receptions of the millimeter-wave radar platform,gesture features are extracted from different angles,and the changes of gesture motion signals at different positions are fully utilized to improve the robustness of gesture recognition.In terms of feature utilization,this paper adopts the multi-angle fusion DTM and gesture motion trajectory RAM as gesture features.In terms of network architecture,the multi-scale feature fusion network proposed in this paper can simultaneously learn the features of palm and fingers.The experimental results show that,compared with other existing gesture recognition algorithms based on millimeter-wave radar,the gesture recognition algorithm based on multi-scale feature fusion proposed in this paper not only takes into account the recognition rate of palm movements,but also significantly improves the recognition rate of subtle finger movements.(2)In view of the low accuracy of gesture recognition when a single feature is input,and the complex network structure and time-consuming calculation when multi-dimensional feature input is used,this paper studies the real-time performance and space complexity of the system,and proposes a lightweight gesture recognition algorithm based on gesture spatiotemporal compression features.Based on the preprocessed RDM,the method uses the dominant velocity of the moving target point to represent the motion feature of the gesture,and realizes the compression of the spatiotemporal feature of the gesture,and then designs a lightweight convolutional neural network to learn and classify gestures.The experimental results show that,compared with other existing gesture recognition algorithms,the lightweight gesture recognition algorithm based on gesture spatiotemporal compression features proposed in this paper not only guarantees the recognition accuracy,but also has great advantages in network model size,system real-time performance and generalization ability.