Research On Gesture Recognition Technology Based On Deep Learning

Gesture interaction is the mainstream way of human-computer interaction,and it is also the research hotspot of many kinds of human-computer interaction.Gesture interaction used to rely on external hardware devices such as digital gloves,EMG,but recently it has gradually developed to use computer vision algorithm to solve the problems in the field of gesture interaction.The gesture interaction method based on computer vision does not need external equipment,which accords with person's interaction habits with computers.The most important problem to be solved is how to improve the accuracy and speed of gesture recognition.In recent years,with the rapid development of deep learning,many problems in computer vision have been well solved.Thus this paper applies deep learning to solving the problem of gesture recognition,and deals with the problems such as gesture recognition in complex environment or disturbed by face and other objects.Some improved methods are also proposed to speed up the algorithms.This paper analyzes a variety of traditional hand gesture segmentation algorithm,aiming at the existing problems of the traditional gesture segmentation algorithm,such as being susceptible to the interference of complex environment,being not robust with high degree of freedom of gesture,easy to deform and other features,then segmentation algorithm of gestures based on convolutional neural network is introduced.Convolutional neural network learns abundant features through network instead of manual design features,so compared with traditional methods,the segmentation accuracy based on full convolution network was greatly improved.In this paper,the structure of the full convolution neural network is further optimized and improved.The structure is compressed into four layers of convolution layer and one layer of deconvolution layer.The real rate of the network model is 93.8%,the false positive rate is 5.3%,and the network runs at 18 milliseconds under GPU,which meets the real-time requirements of human-computer interaction.The whole model can deal with the influence of skin-like objects such as face gesture segmentation.It has high robustness to gestures in different environments,illuminations and angles.In this paper,each module of traditional gesture recognition system is studied.It is pointed out that the results of traditional gesture recognition rely on the accuracy of other modules.The fault tolerance of the whole system is not high,the accuracy rate is low,and it is susceptible to the interference of complex environment,gesture changes and other factors.To solve these problems,this paper designs a gesture recognition algorithm based on object detection network.The algorithm performs intensive prediction to gesture categories and positions on an image at the same time.This paper also improves the network structure,splits the pre-basic network and simplifies it into a small network.After using the network to detect gestures,the tracking algorithm is also used to realize gesture detection based on tracking.By using the improved network,the detection time of a single image is 110 milliseconds,and the accuracy of recognition is 81.2%.On this basis,using the tracking-based detection method,the average detection speed of the algorithm is 33 milliseconds per frame.Experiments show that the algorithm solves the interference of complex environment,face and other body parts' inference on recognition,and has robustness to high degree of freedom and deformation of gestures.It can track and detect gestures in real time,and meets the requirements of real-time gesture interaction.In addition,a set of high-quality,pixel-level dataset of gesture segmentation and gesture detection is also established to train and test the proposed network.