Research Of Static Gesture Recognition Methods Based On Monocular Camera

Gesture recognition based on computer vision is one of the most popular direction in human computer interaction.On the one hand,the traditional gesture recognition method rely heavily on artificial design features,with low accuracy and poor robustness;On the other hand,mature gesture recognition products rely on special devices such as Kinect and Leap Motion,which are expensive.In recent years,deep learning has shined in computer vision,providing a new direction for vision-based gesture recognition.This article mainly studies gesture recognition using monocular camera as the acquisition device and deep convolutional neural network as the main technology basis.The main works of this article includes the following aspects:(1)In order to solve the problem of low accuracy of traditional gesture recognition algorithms,a gesture recognition algorithm based on skin color detection and convolutional neural network is proposed.Firstly,the collected gesture image is segmented by using the skin color detection algorithm combined with YCrCb color space conversion and OTSU.Secondly,the gesture image size is uniformly adjusted to 28 × 28.Finally,the processed image is input to the optimized LeNet-5 model for recognition.For the 9 custom gestures,our algorithm can reach more than 99.61% recognition rate.(2)The traditional gesture detection algorithm has the disadvantages of large calculation volume and slow detection speed.In order to achieve fast and accurate indoor gesture detection,an improved indoor real-time gesture recognition model based on SSD is proposed.Firstly,in order to speed up the detection speed of the model,the basic feature extraction network is replaced by VGG16 to MobileNet.Secondly,in order to improve the detection accuracy of the model,the model was modified as follows: First,the number of convolution layers on the low-level layer of MobileNet is increased to expand the receptive field,the second is to use the feature fusion operation to fuse the low-level and high-level information to make up for the shortcomings of the lack of semantic information of low-level features.The final experimental results show that the proposed model can meet the indoor real-time gesture detection requirements both in detection speed FPS and detection accuracy mAP.(3)Aiming at the problem that the deep learning based gesture recognition has a large number of parameters and cannot be deployed on mobile terminals,a lightweight gesture recognition model Gr-Net based on MobileNet is proposed.First,add L2 regularization constraints to all 3 × 3 deep convolutions to prevent the model from overfitting;second,in order to further compress the calculations and parameters of the model,replace all point-wise convolutions with point-wise group convolutions and add channel shuffle operation.The experimental results show that,under the same hardware platform environment,compared with other lightweight convolutional neural networks,Gr-Net can effectively compress MobileNet without loss of accuracy,and the prediction speed is about 13.39% higher than the fastest model MobileNet-V1,the model size is only 35.4M,which can be deployed on the mobile terminals.