Theoretical And Experimental Study On Gesture Recognition System For3D Scene Video Conference

With the development of information technology, computer is moreand more widely used in human life, and human-computer interaction isgetting more and more attention of researchers. The study for a morepractical human-computer interaction has also become more meaningful.As a common method for human-computer interaction, there have been alot of studies and applications on gesture recognition, but no sufficientresult has been given because of the variety of gesture and the complexityof environment. This article proposed the gesture recognition and controlsystem for3D scene video conference based on Microsoft Kinect, thenimplemented and analyzed this system. The research in this paper haspractical significance.First of all this article introduced the structure of Kinect device, andthe theory for Kinect depth map extraction. Then the method for depthmap extraction and convention to3D point cloud was presented as aconvenience for the following gesture processing tasks. After obtaining3D cloud for the scene, this article started the research from static gesturerecognition. Referring to the process of vision-based gesture recognitiontechnology, the process of gesture recognition needed to locate andextract gesture. By utilizing the advantage that3D point cloud scene hasan additional depth dimension than2D picture scene, this article proposedthe gesture locating method based on distance division, and gestureextraction method based on nearest neighbor theory. After obtaining thegesture point cloud, in the end this article implemented directioncorrection so that gestures would be unified on depth direction. Thengesture feature is extracted by utilizing the depth interval distribution, andis used by support vector machine for model training and recognition experiment. As a result, this article implemented the recognition for digit1to5and achieved a good experiment result.This article also implemented a dynamic gesture recognition partbased on Hidden Markov Models (HMM) for the convenience of videoconference usage. After gesture locating, dynamic gesture feature wasextracted based on the location, velocity, acceleration and direction angleof the gesture moving track. Additional optimization work was also donefor fortifying the stability. This article processed HMM building andrecognition experiment using this dynamic gesture feature, andimplemented the dynamic gesture recognition for4common usedgestures with a good experiment result.This article combined static gesture and dynamic gesture recognitionfor experiment, confirmed the system launch and operation process, andachieved a good implementing effect while decreasing the influence madeby fault gesture operation on the system. The gesture recognition methodbased on Kinect proposed in this article could overcome the problem thatvision-based gesture recognition method is susceptible to environment,achieved a good recognition and real-time effect, and had certain practicalvalue.