Research On Hand Gesture Recognition Model Lightweight Methods Base On Millimeter-Wave Radar

Millimeter-wave radar gesture recognition technology realizes the classification and recognition of gestures by extracting the features of radar echoes reflected by different gestures,which can break through the limitations of gesture recognition technology based on vision and wearable devices.Meanwhile it has the advantages of not being affected by light intensity,good privacy,and non-contact interaction,and has important application value and prospects in intelligent driving,smart home,etc.Gesture recognition lightweight network technology refers to effectively reducing the parameters and calculation amount of gesture recognition network through ways of model compression,lightweight network design,etc.Therefore it can be applied to small devices and embedded devices,and can significantly expand the application scenarios of millimeter wave radar gesture recognition equipment.Millimeter-wave radar gesture recognition network model compression,and lightweight network design are the basic and core elements to realize the lightweight of radar gesture recognition network.In this thesis,theoretical analysis?method research and experimental verification are carried out around the above research contents.The main research contents are as follows:1?Aiming at the problem that it is difficult to accurately extract the echo features of millimeter-wave radar gestures,the multi-dimensional feature map extraction method of gesture echoes is studied,and the gestures are mapped into distance-time,Dopplertime and angle-time maps,which provides the foundation for the recognition network design.2 ? Aiming at the lightweight compression problem of the existing gesture recognition network,the model compression method of multi-dimensional feature fusion gesture recognition network is studied.By performing iterative sparse pruning and knowledge distillation on deep convolutional networks,the amount of parameters and computation of the model can be significantly reduced without affecting the recognition accuracy.3?Aiming at the problem of too many re-training times in model compression,a method of using a lightweight convolutional network to build a gesture recognition network is studied.By adopting the idea of depthwise separable convolution,a good performance is achieved under the condition of lightweight network.4?Aiming at the problem of lightweight network design under different hardware resource constraints,a lightweight gesture recognition network design method based on a slimmable mechanism is studied.By training a shared network with switchable batch normalization layers,the network can adaptively adjust the width parameters to improve the adaptability of the network to the hardware.The validity of the above research content has been verified by theoretical analysis and simulation experiments based on gesture datasets.The results show that the method proposed in this thesis can effectively realize the lightweight of the gesture recognition network on the premise of ensuring the recognition performance.