Some enlightening words in binocular vision (i)

Summarize the words that inspire me:

1, in order to accurately obtain a point in three-dimensional space of the distance, we need to obtain the parameters of focal length F, Parallax D, Camera Center distance TX.

If you also need to obtain the x-coordinate and y-coordinate, then you need to know the coordinate system of the left and right plane and the offset of the stereo coordinate system. CX and CY. where F, Tx, CX and CY can obtain the initial value through stereo calibration, and through the stereo calibration optimization, so that two cameras are mathematically completely parallel placement, and left and right camera CX, CY and f the same ( This is the ideal form for fully parallel alignment of the left and right views in Figure 2). The work of stereo matching is to find the last variable based on the previous one: Parallax D (this d generally needs to achieve subpixel accuracy). This completes the preparation required to achieve a point-three-dimensional coordinate.

With that in clear, we know that all of these steps: calibration, calibration, and matching are all centered around how to get more accurate F, D, Tx, Cx , and Cy .

2, the world coordinates the origin point is the left camera convex lens light heart.

3, in the pinhole model, the light passes through the pinhole (that is, the lens Center) at the focal length of the image, so the image of Figure 3 is the camera's CCD sensor surface. Each CCD sensor has a certain size and a certain resolution, which determines the conversion relationship between millimeters and pixel points. For example, the size of the CCD is 8mm X 6mm, the resolution is 640x480, then the conversion between millimeters and Pixels is 80pixel/mm.

In practice, we mathematically equate this image plane with a small hole, which is equivalent to assuming a virtual CCD sensor before the center point of the lens.

4, this has been mentioned in the previous article. In the actual measurement, due to the camera placement of the relationship between the left and right camera F, CX, Cy are not the same. In order to make the left and right view to achieve a complete parallel alignment of the ideal form so as to achieve the convenience of mathematical computation, the work of the stereo calibration is in fact, in the case of the largest overlap area, let two camera axis of the forward parallel, and let left and right camera F, CX, Cy the same. Therefore, the values in the Q matrix are different from the values of the two instrinsic matrices, which can be understood.