The Research On Static Gesture Recognition Based On Kinect Depth Data

With the development of information technology, human-computer interaction technology has gradually become an important part of our life. As an important research field of human-computer interaction, gesture recognition has become a research hot-spot in recent years. Traditional gesture recognition methods are mainly based on the mouse and pen, visual and data glove, and they mainly include template matching, neural network and statistical analysis, etc. But these techniques and methods have some limitations in applications. The method based on vision is susceptible to light and complex background conditions, and its recognition rate is low. The method based on data glove need special sensing devices, and users are very inconvenient. In 2010, Microsoft invented Kinect somatosensory devices. The device can extract depth data of the gesture, and overcome the interference of light and complex background.In this thesis we analyze the current research status of gesture recognition, and find that the effect of the methods based on vision is poor. In addition, most of the methods of gesture recognition don't have rotation, scaling and translation invariance. To overcome the above shortcomings, we use the Kinect sensor and SURF(Speeded Up Robust Feature) algorithm for gesture recognition, and propose a static gesture recognition method based on depth information of Kinect sensor. Thesis mainly studies the following several work:1. Gesture segmentation based on depth data. First of all, Kinect sensor collects the depth data of the users and the background, and extracts the index value from these depth data. According to these different index values, we segment the human body from the background. Finally, we segment the gesture from human body by the threshold value.2. Feature extraction based on the improved SURF algorithm. SURF algorithm mainly includes and feature description. According to the characteristics of the binary image, this thesis improves the method of the key point detection and the dominant direction calculation.3. Gesture classification based on Bayesian regularization BP neural network. Most of the training methods of BP neural network are easy to lead to over-fitting. In order to overcome this defect, this thesis uses Bayesian regularization BP neural network as the method of gesture classification. The experiment shows that the generalization ability of the method is ideal.4. Gesture classification based on decision level fusion. This thesis proposes a method of gesture classification based on multiple classifier decision level fusion. This thesis uses the maximal finger angle as the gesture features, and builds a new classifier. We analyze the classification results of the method and classification results of Bayesian regularization BP neural network, and obtain the final results.Finally, this thesis analyzes the proposed gesture recognition method by the experiment, and the results show that the method is feasible. Through experimental analysis, this method overcomes the interference of light and complex background and so on. So the method based on Kinect depth information overcomes the drawback of the method based on visual, and provides a new research direction for gesture recognition technology. This method can be extended to many kinds of gestures, and be used in sign language recognition, etc. It not only has the very high application value, but also laid a good theoretical basis for further study.