Research On Micro-Doppler Based Hand Gesture Recognition Method Using Deep Learning

In recent years,as a typical representative of human-computer interaction,gesture recognition technology has very broad application prospects in the fields of augmented reality,autonomous driving,smart home,etc.because of its touchless,naturalness and flexibility.More mature technologies of hand gesture recognition are based on visual sensors and wearable devices.However,both of these two methods have certain limitations.Visual sensors are greatly affected by lighting conditions,and wearable devices have a strong sense of presence,which seriously affects the user’s experience.The radar-based gesture recognition technology is not affected by lighting conditions,and does not require users to wear additional sensors,has the characteristics of high spatial resolution,low energy consumption,and easy integration.This thesis focuses on the micro-Doppler based gesture recognition method,the main research content is as follows:(1)Aiming at the insufficient utilization of micro-Doppler features of gesture data,a gesture recognition method based on a multi-scale features fusion network is proposed.This method uses pyramid convolution to increase the width of the model,and to enhance the ability of extracting multi-scale features in the same layer.This method promotes the information interaction between feature channels by using 1×1convolution,and combines low-level visual features and high-level semantic features by adding short connections.As a result,the effect of gesture recognition is significantly improved.(2)Aiming at the problem of large size and slow running speed of ordinary models,a gesture recognition method based on lightweight networks is proposed.This method improves the MobileNetV2 model by introducing dilated convolution,the early downsampling strategy,and Hard-Swish activation function,which can improve the convergence speed and stability of the model.This method expands the receptive field and makes full use of context information without increasing parameters and floating point operations,also quickly reduces the resolution of the feature map,parameters and floating point operations without losing too much detailed information.As a result,the accuracy and speed of gesture recognition are well balanced.The comparative experiments on the radar gesture data set show that the proposed method can complete the dynamic gesture recognition task more effectively,and distinguish the gesture categories with higher similarity more accurately.Among them,the overall accuracy of the gesture recognition method based on the multi-scale features fusion network is as high as 98.63%,and the gesture recognition method based on the lightweight network only uses about 2.22 million parameters and 249 MB of floating point operations to reach 96.51% overall accuracy.At the same time,this thesis also performs generalization performance analysis on data sets with different data volumes,noise levels,short-time Fourier transform window lengths,different radar aspects,and unknown subjects.Experimental result shows that the gesture recognition method based on multi-scale features fusion network and the gesture recognition method based on lightweight network both have good generalization performance on the radar measured data set under various conditions.