Architecture for 3D Human Motion Tracking

This week, I also learned about a new framework for 3D Motion Tracking. This framework uses random sampling and local optimization to achieve a better compromise between robustness and effectiveness, A local optimization method based on simulated human motion is introduced to improve the superiority of tracking.

The framework uses the cylindrical modeling method (see figure 1) to model the three-dimensional human body. Except for the non-standard trunk cylindrical (the cross section is elliptical), the rest parts are standard cylindrical (Head, forearm, hind arm, thigh, calf), and number part. The origin of the local coordinate system of the non-trunk part is located in the center of rotation of each cylindrical base, and the origin of the whole body is located in the center of the trunk. At the same time, degrees of freedom are introduced to constrain the movement of each part. Both the forearm and the calf have a degree of freedom. The first, backarm, and thigh have three degrees of freedom, and the trunk has six degrees of freedom. In this way, 25 degrees of freedom is used.

Figure 1 3D Human Model

Then perform 3D Human Body reconstruction. Here we use Visual Hull Method to reconstruct 3D scene points and indicate normal vectors. The reconstruction points are obtained through the intersection of the observed cone, the corresponding normal is obtained through the cross product of the line observed and the tangent of the image contour.

The local optimization method is an extension of the iterative closest point algorithm. It simulates human motion based on the registration algorithm. The advantages of this method are as follows: A simple and effective location update solution for hierarchical model is proposed to replace recursion to solve a series of dynamic equations.

Experiments show that this method is more robust and effective, and it is also better than the method based on image force and particle filter, particle Filter requires many particles to achieve a certain effect, so this method is slow. The image force-based method is unreliable and easy to lose. It is only suitable for simple motion such as walking, but it is unreliable for boxing.

This method has a major limitation: You need to use similar motion models for detailed and patient training.