Research Of The Dynamic Gesture Recognition Based On The Kinect Depth Image

Gestures as a very important way of human-computer interaction, through the computer, we can detect the gesture and tracking, recognizing to understand people's intentions from the video, due to its natural and convenient, more suitable for human natural interaction requirements, its application scenario is very broad, it has become one of hottest topics in the study of human-computer interaction.Traditional gesture recognition technology mainly includes the gesture recognition based on wear equipment and gesture recognition based on the ordinary camera RGB computer vision. Gesture recognition based on wear equipment refers to the use of data glove and three-dimensional devices for gesture recognition method, it limits the natural human-computer interaction. Gesture recognition based on vision research contains many algorithms, but these methods in the gesture at the time of the partitioning easily affected by the factors such as illumination, background and camera features, it can cause the recognition rate is not high, poor robustness. So the traditional gesture recognition technology don't have a good application prospect, this thesis studies the dynamic gesture tracking algorithm based on Kinect, can well solve these problems.Based on the predecessors'summary and add some research based on Kinect, This thesis studies the two ways of existing dynamic gesture recognition, they respectively are dynamic gesture recognition based on curve fitting and dynamic gesture recognition of making use of monotonicity of track. Dynamic gesture recognition based on curve fitting is mainly aimed at the curve fitting according to the center of gesture extracted, after the fitting using Hu curve moment for training and classification, this algorithm has the advantage of high calculation efficiency, but the recognition rate is low. Dynamic gesture recognition using the monotonicity of trajectory by each different monotonicity of each path to distinguish different track meaning, the characteristics of this algorithm is that if the curve monotonicity is not frequent change, it will have a good recognition effect, but if the curve monotonicity changes are frequent,the recognition rate is not very good. Through summarizing the advantages and disadvantages of the above two methods, this thesis put forward the dynamic gesture recognition based on direction vector quantization method. this method adopting the Camshift algorithm to do the gestures centroid extraction, after the gestures centroid extraction using cubic spline curve to do the centroid correction, using DTW matching algorithm to do the gesture trajectory classification and identification, the advantages of this algorithm is that it can to solve the gesture changes too fast and we will not be able to accurately track and for the same gesture, it can cause different trajectory by the different people it cannot accurately identify, the two problems by using this method could be well resolved. For these three algorithm, this thesis made a series of experimental studies, this thesis finally concluded three algorithm of recognition rate, good or bad robustness and a series of conclusions, finally this thesis come to the conclusion that the method this thesis come out is more efficient.