Research And Implementation Of Key Technology For Dynamic Hand Gesture Recognition

Dynamic hand gesture recognition is a crucial research topic in the field of human-computer interaction which has both important theoretical significance and broad application prospects. Under the overall framework of Kinect based dynamic hand gesture recognition system, gesture segmentation, gesture tracking and dynamic hand gesture recognition are investigated in this thesis.In gesture segmentation, if only process the Kinect color images, hand gesture segmentation is susceptible to interference factors such as light, skin color objects and so on. However, only processing the Kinect depth images is easy to contain the area besides hand gesture. To solve the problems, a novel gesture segmentation method is proposed based on the depth information and the local skin color detection. As traversing the whole image to detect the skin points is time-consuming during the process of skin color segmentation, the local skin color detection approach is applied to improve the real-time performance of this system. Experimental results demonstrate that the proposed method can accurately segment the target gesture from the background.In the dynamic hand gesture tracking process, an improved TLD(Tracking Learning Detecting) tracking algorithm is proposed for the current TLD method is prone to cause drift problem when the target gesture is occluded. In the situation where the target gesture is seriously blocked, the Kalman filter and the Markov direction predictor is added respectively in the tracker and detector of TLD to predict the position of the target gesture in the current frame, and as a result can enhance the ability to distinguish the similar gesture trajectory. The experimental results show that the improved TLD tracking algorithm can effectively solve the target gesture occlusion problem.After completing the feature extraction of hand gesture by a local gesture feature extraction method, in order to simplify the procedures and improve the computation accuracy of the HMM(Hidden Markov Model) based B parameter train process, an improved HMM dynamic gesture recognition algorithm is proposed to predict the B parameter based on SVM(Support Vector Machine). Through the recognition experiments on gesture trajectory features of 0~9, the validity of this improved dynamic hand gesture recognition algorithm is verified.Finally, an intelligent wheelchair based hand gesture recognition system is designed, and 5 corresponding control instructions are defined to control the movement of the intelligent wheelchair. Experiments conducted under the laboratory environment for multiple times demonstrate the high real-time performance and stability of this system.