3D Hand Pose Midiling And Interaction System Research

3D gesture with its convenience and flexibility features, has great value in HCI application fields. The core technology of 3D gesture is the precise hand posture estimation and analysis. This paper focuses on 3D hand gesture estimation techniques in interactive scenarios. With the purpose of hand pose that estimated is under topological restriction by an efficient algorithm, we develop a hand pose estimation framework which combines the palm pose estimation and finger pose recognition. Moreover, we design a HCI application based proposed 3D hand pose estimation algorithm. To conclude, the research content of this thesis are as follows:Firstly, we carried out the research on 3D hand posture estimation and molding methods in interactive scenarios and summarized the main drawbacks of them. Specifically, to take the global hand topological restriction and both the efficiency and accuracy of our algorithm into consideration, based on the current cascade regression algorithm, we designed a hand pose estimation algorithm framework which from palm posture estimation to finger pose recognition, which is efficient and the result of hand pose under topological restriction.Secondly, to detect the hand palm pose more faster, this paper presents a 3D fingertip detection based on depth data, by means of 3D finger direction detection and the 3D palm direction estimation. Compared with the traditional PCA estimation by test on the public dataset, the result of 3D finger direction by our method is similar with result of the PCA method. Under the result of 3D palm direction, the result of our method is 33% lower than the PCA method. This result show that our proposed method got a much better performance on hand global pose estimation.Thirdly, we treat the finger pose as parameter of skeleton node, and use the random regression forest model with 3D Pose-Indexed feature to predict finger parameter, In order to improve the accuracy of finger posture estimation, we induct a methodology by using random forest to adjust hand posture based on the joint features which is computed via depth image and joint positon. Compare to cascaded pose estimation algorithm, our methodology is more efficient and the result of hand pose is under topological restriction. The testing result in MSRAH15 databases shows that per joint error of our method is lower 22% than the PCA themod, The per joint error can also be reduced about 3% by finger pose correcting. The result in ICVL dasets shows that per joint error is 17 mm under our method, witch is a little worse than other algorithm, but is more efficient and has more promissing pose result. In the end we design a HCI application which is pretty example work in feature via proposed hand pose estimation algorithm.