Research On Gesture Recognition Technology Based On 3D Separable CNN

The existing interaction modes of Augmented Reality(AR)products mainly include gesture interactions,voice interactions,and gaze interactions,among which dynamic gesture interactions are one of the most natural interaction modes.At present,the research of dynamic gesture recognition is mainly based on large sample size.The amount of model parameters and calculation is large,and the requirements for hardware are high,so it is difficult to apply it to mobile devices with limited memory.In this paper,the lightweight model of dynamic gesture recognition is studied to further improve augmented reality gesture interaction system.In the dynamic gesture segmentation section,two types of common gesture segmentation methods are used to perform experiments under the conditions of different lighting conditions and different background images,the feature extraction performance conditions of different methods are explored.According to the advantages of HSV method in quickly extracting gesture regions and filtering background information and the advantages of motion history image method in extracting gesture motion information,the HSV and motion history image method is adopted as the dynamic gesture segmentation scheme of this paper.In the dynamic gesture recognition part,a three-dimensional separable convolutional neural network is proposed to realize dynamic gesture recognition and network model compression.The convolution process of three-dimensional convolutional neural network is separated to compress network parameters and increase the speed of calculation.Skip connection and layer-wise learning rate are used to solve the problem that the separated network is difficult to train.At the same time,the feature diagram is promoted through the shuffle operation.A gesture data set is established for dynamic gesture recognition of AR glasses,and the deep network model is used to visually analyze and fit the sample so as to simplify the complex nonlinear distribution problem.In the part of dynamic gesture interaction system,the three-dimensional separable convolutional neural network model was applied to Hololens holographic glasses with Unity 3D,and the feasibility of the method for realizing dynamic gesture recognition on Hololens holographic glasses was verified by testing the target gesture.