1

I am currently working on a head motion tracking project using normal webcam. Face detection and LK tracking works fine but hope to build something more comprehensive. Is it possible to use the coordinates of the feature points in one frame to the next to estimate the 6 DOF (scaling, rotation) motion of a subject? Intuitively I think it is possible because we can assume that the feature points on all the frames fall on the same plane. I am looking to measure the pitch yaw and roll and scaling of the subject. Scaling is easy if we assume that the subject only experience translation (no pitch yaw roll) by just looking at the different in distance between any two given feature points. Pitch yaw roll and scaling becomes tricky when they all combine. I am trying to find out how Kinect capture motion because my guess is I can treat the feature points like the infra red pattern for Kinect.

Please advice on this. Need to find out the feasibility of this method before I try something else.

THanks, Kelvin

4

1 回答 1

0

不能完全建议您提出的可行性，因为这取决于您的特征提取方法，即您如何获得特征点？如果您的拍摄对象总是面向相机，那么您可以应用主动外观模型进行头部姿势估计，这将告诉您统计计算的 3d 面部网格的位置适合新的面部图像。一个叫 Pedro Martins 的人在这方面做了很多工作，包括 AAM+POSIT 跟踪和头部姿势估计http://www2.isr.uc.pt/~pedromartins/我自己也使用 AAM 来提取面部表情，作为一个问题，它需要比 6 DOF 更高的个人面部特征跟踪精度。对于 AAM，除了跟踪应用文献外，还有很多开源实现，包括 Pedro Martins 自己的姿态估计实现。

然而，正如我在上面写的那样，我意识到你可能没有头部朝向相机的情况，因此减少了 AAM 的使用。您的数据看起来如何？你有示例视频吗？

PS 至于 Kinect——我不知道你项目的情况，但 Kinect 2 会有足够的分辨率来跟踪全脸网格，但它会在明年发布。一般来说，我希望有人可以对旋转/缩放估计的头部跟踪问题给出更有经验的答案。

于 2013-10-16T14:54:32.723 回答