A Study Of Gesture Recognition Based On Deep Learning

In 1936, Professor Alan Turing proposed a new and abstract computing model---Turing Machine. Modern computer based on Turing machine has been improved a nd developed rapidly and efficiently. Now computer has been the helper and equipment which is absolutely necessary in people’s daily life, working and study. Computer can be found in every corner and detail of human society. With the develop of computer, p eople also seek a simpler, more direct and more natural human-computer interaction.Among the existing human-computer interaction methods,the one based on gesture could be a natural, intuitive, effective and relatively simple way that adapts to human habit. Compared with traditional human-computer interaction method such as mouse and keyboard, the gesture method could more adapt to human habit. Compared with the posture method, the gesture method has a more uniform standard which could be easy to distinguish. Compared with the speech method, the gesture method is much simpler whose recognition rate is much better. Due to the reasons above, gesture recognition and the human-computer interaction based on it can be used in many fields, such as motion sensing game, remote control and the scene that ap hasiac or deaf people interaction with machine and other people smoothly. Therefore gesture reco gnition and the human-computer interaction based on it have got tons of attention from people.There are two main ways in the gesture recognition, one based on wearable devices and the other based on vision recognition. The method based on wearable devices catching information from hands has high recognition rate and low computational amount. However the devices are unwieldy and expensive. The method based on vision recognition susceptible to internal algorithm generally has the disadvantage that the recognition rate is not enough good.To overcome this difficulty, the author connects deep learning algorithm with kinect and improves the deep belief network model. K inect developed by Microsoft has the reasonable price and the function that captures both RGB images and depth maps. Deep belief network proposed by Hinton is such a deep learning algorithm whose property has been proved in many tasks. The creative double deep belief network in this thesis isperformed to train the RGB images and depth maps caught by kinect and fulfil the gesture recognition tasks.Comparison experiments show that the innovative method proposed in this thesis has much better results.