联合LBS和Snake的3D人体外形和运动跟踪方法

3D Human Body Shape and Motion Tracking by LBS and Snake

为了解决基于多目视频轮廓信息的3D人体外形和运动跟踪问题,提出一种联合线性混合蒙皮和Snake变形模型的算法框架.首先建立人物对象的蒙皮模型,以每一帧多目同步视频的轮廓作为输入,采用一种基于剪影轮廓的可视外壳重建算法,使得作为3D特征的可视外壳保持了局部细节且更加光滑;并使用关节型迭代最近点算法进行匹配以捕获出每一帧骨架子空间下的人物3D外形及运动;再一次使用当前帧的多目轮廓信息,让Snake内外力共同作用于人物网格模型上的顶点,使之自由地趋近于目标对象.使用带ground-truth的合成数据进行对比实验的结果表明,该方法因同时使用3D误差约束和2D误差约束,提高了跟踪精度.

A new framework combining Snake deformable model with LBS（Linear Blend Skinning） is proposed for both 3D human body shape and motion tracking.First,skinned mesh model of object is established.Then surface mesh with highly improved quality is obtained by using our new silhouette-based visual hull reconstruction method for each frame of multi-view videos.After that,articulated ICP registration method is used to capture the 3D human body shape and motion under Skeletal Subspace Deformation.Furthermore,the multi-view silhouettes are applied again for Snake external force computation,and then the vertices of surface mesh are moved to approach the target object under silhouette and internal forces.Synthetic data with ground-truth is used for quantitative comparison,and experimental results show that the tracking accuracy is improved since both 3D and 2D error constraints are used.