计算机辅助体训中基于形变度与运动分解的运动员3-D形态与运动信息识别

在计算机辅助体训中根据所获取的动态图像序列研究3-D空间中运动员的形态与运动信息识别问题。首先，为了有效控制识别过程中所需加入的约束程度，将根据运动员图像序列研究不同运动项目中运动员的形变度。这一形变度计算应能克服图像序列中特征点丢失以及方向性噪声等不利因素。为了有效地识别形态与运动参数，我们将利用小波变换的方法并结合形变度对运动员非刚性运动进行运动分解。其中形变度用于控制运动分解的规模。这种运动分解的目标在于运动基可以事先确定，同时使参数求解能与传统的刚性分析方法得到统一。在随机场的非整则邻域系中，我们还要通过研究克服传统确定方法仅仅利用图像平面的欧氏距离所带来的无法有效反映运动局部平滑性和相似性的缺陷，通过神经计算的方法确定邻域系从而有效地体现运动员的局部运动细节并提高求解的鲁棒性。该研究的完成将能对3-D空间中体育运动员的形态与运动信息进行鲁棒识别，并在各类计算机辅助体训中得到应用。

Recognition of 3-D structure and motion for athletes will be investigated based on dynamic image sequences for computer aided sport training. First, in order to control the constraint degree in the process of recognition, the estimation of deformation degree of athlete non-rigid motion based on image sequence will be analyzed. For real image sequence, we should overcome the negative factors such as directional uncertainty and missing feature points. Meanwhile, we will factorize the athlete motion using deformation-degree-based wavelet transformation. The factorization scale will be controlled by the deformation degree. The objectives of this new factorization approach lie in that the motion bases can be determined beforehand and the non-rigid motion analysis approach in this project is unified with the traditional rigid methods. Traditional methods for the definition of irregular neighborhood system of random field use the information of Euclidean distance of feature points in image plane. Therefore, the motion smoothness and similarity of local motion can not be effectively reflected. To overcome this limitation, we will investigate to define the irregular neighborhood system based on neuro-computing. Thus, the reflection of local motion the robustness of the motion estimation will be improved. We can get robust recognition results for athlete structure and motion when this project is finished. The achievements of this project can be applied to real computer aided sport training areas.