Gesture Recognition Based On Deep Learning

With the rapid development of science and technology, human-computer interaction is more frequent occurred around more people, bringing convenience to the lives of many people, where human-computer gesture interaction is the most active area. However, the hand is non-rigid objects, rich variations, Increasing the difficulty of the gesture recognition. Deep Learning has made great progress since 2006, it has brought new hope to artificial intelligence, which showing excellent performance gives us enough confidence to complete the gesture recognition task. Gesture recognition is a complex system, usually gesture recognition system based on computer vision includes gesture definitions, image acquisition, processing, analysis and understanding. A good identification system is often inseparable from a good pre-processing, but this does not mean that the pre-processing of the image is the most important. In my paper, I have focused on the study of gesture recognition algorithm.In the implementation process of gesture recognition, this paper studied from the two main aspects of deep learning, one is based on the Restricted Boltzmann Machine(RBM), and the other is based on the LeNet-5 Convolutional neural network(CNN). This paper also presents a combined network CNN and RBM: using deep stacked network formed by RBM for unsupervised feature extraction, combined with supervised feature extraction of CNN, and finally using integration of these two types of features for classification. Experimental results show that our proposed combined networks perform better in pure background gesture samples and recognition capabilities for complex background gesture samples are to be improved. In addition, compared to the traditional three-layer neural network DL has obvious advantages, while DL greatly improved the training of multi-layer network problems, from our experiment can be seen the performance of multi-DL is better than traditional networks.For the gesture control of unmanned aerial vehicle, defining ten categories basic gestures first, while the samples are divided into two part, one is under simple background and the other is under complex background, and make a simple pre-processing for these samples. Using five networks for a static image recognition, which are Deep Belief Networks(DBNs), Deep Neural Network(DNN), CNN, combined Network of CNN and RBM, and the traditional three-layer neural network. For the gesture recognition in video stream, this paper is divided into three part: detection, tracking, recognition, and making a detailed description. The results show that CNN can well adapt to gesture detection, tracking and recognition under complex environments.