Human 2.5D Thinning Algorithm And Its Skeleton Extraction Research

Human action recognition has a variety of applications in gaming, human computer interaction and even healthcare domain. Gesture recognition offers more intuitive and convenient control than traditional input device in human conputer interaction. Centring on the human pose estimation, we mainly study the human skeleton estraction algorithm and put forward 2.5D thinning algorithm. Finally, we build a multimedia interactive system for exhibition hall based on the human skeleton information.The main contribution of this thesis is in the following:To overcome the drawback that traditional thinning algorithm cannot handle the depth information in depth image, we put forward a 2.5D thinning algorithm. This algorithm can make full use of the distance information offered by depth camera to automatically segment arms from body region and generate multi-layer thinning lines.Based on the multi-layer thinning lines extracted by 2.5D thinning algorithm, we put forward a human 3D skeleton extraction algorithm. We detect the anatomical landmarks of the human skeleton through endpoints and line distance information to overcome the self-occlusion problem. Then we build a standard human model and use inverse kinematic method to match it with landmarks. The model fitting method ensures that the length of joints satisfy the standard scale. The algorithm runs at 14 frames per second and the average error is 5.89 pixels. Comparing with the traditional unsupervised method which runs at about 2-4 frames per second, it improves the speed a lot. It can extract the skeletons from more poses than traditional supervised method.Finally, we use the imformation of human skeleton to build a multimedia interactive system. This system can recognize gesture by judging the location and speed of the joints in human skeleton. It can control the exhibition content with waving, clicking and pointing in air. The waving recognition rate can reach 97.25% and the pointing precision is 2.15 ?. This system simulates the page turning action of human to make the gesture more natural.