Research On Vehicle Scene Gesture Recognition Based On Deep Learning

As an intuitive and natural way of interaction,gesture has gradually developed from daily interpersonal communication to new human-computer interaction,especially in vehicle control system.Many high-end vehicles have applied gesture recognition technology to vehicle audio-visual entertainment system.In many technology implementation routes,the machine vision scheme based on monocular color camera has been paid more and more attention by researchers because of its simple and cheap sensors.In order to solve the recognition difficulties such as complex background,self occlusion and easy deformation,this paper uses deep learning technology to improve the recognition accuracy.In order to solve the above problems,this paper designs a deep learning gesture recognition solution based on monocular color image.The scheme consists of three parts:Firstly,the hand segmentation algorithm is used in the complex background of vehicle scene.For the problem that the existing hand gesture segmentation method is difficult to segment static gesture efficiently and accurately from the image in complex background,an improved background difference method is proposed,which is used to detect the dynamic target trajectory.That is,through background modeling,the image and background model are differentiated on a specific threshold,and finally,the gesture image is segmented by the connected domain and morphological processing.Compared with the previous segmentation method based on skin color and contour,the algorithm has better segmentation effect and stronger robustness through the verification of segmentation effect in different scenes.Secondly,three classical classification convolution network models are compared in detail.Based on the previous gesture segmentation algorithm,the model configuration and classifier training method are selected.VGG16,Res Net50 and Inception V1 are used to perform recognition experiments on ASL authoritative gesture data set,and the experimental results are compared.The experimental results show that only using color images,using the classic high computational power convolutional neural network can achieve good recognition effect,far better than the traditional machine learning method,and Res Net network structure is the most suitable for gesture recognition task,with the highest accuracy of 87.3%.Finally,the network model is lightweight modified for mobile deployment.For the problem that the classical CNN network model has large parameters and is difficult to deploy to mobile terminal,this paper first introduces the network structure and lightweight principle of Mobile Net,and tests the performance of gesture recognition task on ASL data set.Finally,a new lightweight gesture recognition model based on Mobile Net network is proposed.The residual modules can be separated and convoluted in the network to reduce the loss of information layer by layer,and the feature information of the volume layer before and after communication can be improved to improve the final recognition accuracy.These connections include simple constant connections with output and input dimensions,and jump connections with 1 * 1 convolution that are different from output dimensions.The experimental results show that the new network structure is suitable for gesture recognition task.Compared with Mobile Net,the accuracy of the new network is improved by 1.7 percentage points when the computation and parameters are constant.