Abstract: While the research on 3D human motion reconstruction from video has progressed rapidly, but most approaches rely on accurate foreground silhouettes, which lose the inner texture information. This paper presents a novel method of reconstructing 3D human motion using image texture, and proposes a robust and adaptive tracker model. The model is based on L1 tracker, and can extend L1 tracker to the case of articulated body parts, in multiple view images with hierarchical searching. The tracker model can find the target with minimum reconstruction error from the template subspace, and this error in every view is taken as the likelihood measure function. The method is performed in annealed particle filter framework. To improve the tracking performance, two special methods are employed in the texture template updating process. First, 3D human body model is used to predict self-occlusion which is more accurate than image texture. Secondly, the template coefficients are used to detect bad target found by the tracker model. Combining these two detectors can avoid bringing an improper tracking result to texture template set. The proposed method shows good performance on HumanEva-II dataset.

Key words: 3D human motion reconstruction, L1 tracker, multiple view, hierarchical search, template update

摘要： 近年来从视频中恢复三维人体运动的研究发展很快，其中大部分方法是基于前景轮廓的。提出了一种基于纹理信息的三维人体运动恢复方法，并给出了一个鲁棒、自适应的跟踪器模型。该模型基于L1跟踪器，并将其扩展到多个视图中，使用分层搜索来跟踪人体的各个部位。它可以寻找在模板子空间里重构误差最小的跟踪目标，将每个视图的重构误差作为衡量人体三维姿态与图像拟合的可能性函数。整个算法在退火粒子滤波的框架下进行。为了提高跟踪准度，在纹理模板更新过程中使用了两种方法：用人体的三维模型来检测自遮挡；根据模板系数检测计算错误的跟踪结果。综合这两种检测器，可以防止遮挡后和计算错误的跟踪结果加入到纹理模板中。在HumanEva-II测试集上的实验表明，该算法能够得到较好的结果。

关键词: 三维人体运动恢复, L1跟踪器, 多视图, 分层搜索, 模板更新