How to Learn about gesture tracking: Realtime and robust hand tracking from depth (3) Cost Function

Iker original, reproduced please indicate the source: http://blog.csdn.net/ikerpeng/article/details/39050619

Realtime and robust hand tracking from cost function learning in Depth

First, we should know what the input data is: 3D point cloud data.

3D point clouds give me this feeling.

The output is a hand-fitting model (48 sphere model ).

The 3D point cloud data here is represented by P, and each sphere is represented by SX. The center of the I-th sphere of ci; D indicates the depth graph (distinguishes another D (.)). Let's take a look at the following cost function.

Understanding of this formula.

In short, it is the difference evaluation, and ultimately it needs to be minimized.

First, let's look at the first item: its definition is as follows:

Here, D represents distance, that is, to calculate the distance between point cloud P and center C of the sphere and then subtract the radius. That is, the distance between the point cloud point and the sphere surface. P obtains a subset of sub (p) = 256 point clouds. This is designed to reduce the complexity of computing and ensure accuracy.

The calculation amount here is: the distance between the 256 points and the center calculation of the 48 sphere respectively (in fact, it is a number of Sphere SX (p) closest to point cloud p )). Finding the hand parameter that gets the minimum distance is the solution of this equation. However, this is a problem. For example, a sphere with a small distance to a point cloud may be at any position around the ball center. Therefore, we need to modify it. Introduce the second item:

The purpose of this item is to enable the sphere to be In the vertex cloud. Where J (c) represents the projection point of the sphere on the depth graph. If the depth D (J (c) at J (c) is very close to the zcoordinate of the center of the sphere, then the center of the sphere is above the depth, add the absolute value of their difference to the penalty factor. If the depth at J (c) is not, it indicates that the sphere center is outside the depth graph contour and J (c) is added) the distance to the depth graph contour is used as the penalty item. (Note: What I understand here is not very profound. I don't know how to calculate the distance from J (c) to the depth graph contour)

Finally, we need to restrict the relationship between the ball and the ball. That is to say, the ideal condition between the ball and the ball is tangent, and too far and too close is not very good, so we need to add a penalty factor.

The two balls on the left are not suitable, and the two on the right basically meet the requirements. Therefore, the third amendment is introduced:

The significance of the formula is very obvious.

Let's take a look at the quantitative factor as a whole, because the first item and the second item have the same effect. A simple quantitative processing.

 

 

Problem:

1. How much original 3D point cloud data exists? Here we use it to fit it. Does it mean that the hand detection has been achieved?

Http://blog.csdn.net/opensource07/article/details/7804246 is the data generated by a point cloud.

2. How is the initialization of the 48 golfers carried out? (Implementation of finger tracking ?)

3. Understanding of the second amendment?

This section is complete.

How to Learn about gesture tracking: Realtime and robust hand tracking from depth (3) Cost Function