Research And Application Of Gesture Recognition Based On Deep Convolutional Neural Network

Various smart products are constantly emerging under the background of intelligent technology,such as: smart homes,smart factories,intelligent transportation,etc.The way of human-computer interaction continues to innovate,from keyboard and mouse to touch screen,even to face recognition and gesture recognition.Therefore,the human-computer interaction approach is moving toward intelligent and non-contact trend.Contactless human-computer interaction based on gesture recognition has always been a research hotspot and has produced quite good results from wearable gesture recognition to traditional computer vision,because the hand is the most flexible and controllable organ of the human body and can convey different information according to different postures in human-computer interaction.While,the wearable gesture recognition is expensive,and traditional computer vision gesture recognition is not robust and suitable to complex environments which relies too much on expert experience to take "feature engineering".The deep convolutional neural networks are widely used in recent years which have strong abilities of feature extraction and model generalization and can reduce manual intervention with set feature extraction and classification in one,so it can construct gesture recognition system quickly and easily.In this paper,the deep convolutional neural network model is used to realize offline real-time gesture recognition on the mobile end.The main research contents are as follows:Firstly,the effects have been analyzed and compared of the gesture feature extraction of traditional machine vision and deep convolutional neural network under different environments.Experimental results show that the deep convolutional neural network is more robust in feature extraction.Then,a network model has been designed that can be used for gesture recognition with analyzing several efficient deep convolutional neural network models.Next,collecting data sets and using Tensorflow to write model and trainning program to train the model.And the developed Android apps using the trained model is applied on the mobile phones.Finally,the recognition performance of the model is tested.According to the test results of the mobile terminal,the model designed in this paper has an accuracy rate of more than70%,except for gestures with symbols 2,3 and 4,on the premise of a small number of parameters and calculation.And the ability of contactless human-computer interaction istested on a balance car,which developed in this paper.