Panorama picture Stitching

Recently in doing a teacher cloth homework, the original is to do panoramic video synthesis, but made to find real-time nature is not up to, change into a panorama image synthesis bar. Don't say much nonsense, talk about the principle.

Principle:

1 histogram equalization The image three channels are normalized.

2 Cylindrical projection

3 Picture Registration

4 Image Fusion

Simply explained that the cylindrical projection is not difficult, is a curved surface and a tangent to the triangular transformation of the relationship between the people who have high school mathematics can fully understand. To do this program, you need to know the camera's angle of view, or know the distance between the projection center to the projection cylinder, these two parameters can be converted to each other, know one can.

Picture registration, because of the time, I use the algorithm is the simplest, the first is to extract feature points, I use the SIFT feature points, this feature point calculation is really slow ah, this is no way to achieve real-time nature of the main reason. Picture after projection, horizontal, vertical translation, registration.

For image fusion, it is weighted fusion. In the area of two images coincident, dst=a*src1+ (1-a) *src2,src1 is a picture, Src2 is the second picture, DST is the target image of the fusion. Where a is constantly changing, the aim is to achieve an effect, in the near Src1, a relatively large, so that the proportion of SRC1 is relatively large. As it gets closer to the SRC2, a slowly becomes 0, so that the second picture takes up more and more weight.

Here are some of the problems you're experiencing:

1 different pictures, the number of SIFT feature points is not the same, so for different pictures, parameters may be set differently.

2 Sift feature point acquisition speed is too slow, I use the OPENCV library function to get the feature points, do not know is not OPENCV library function problem. Light get feature point, a picture is almost 1s, not up to the real-time requirements.

3 According to the image fusion of how much, you can probably judge the distance of the feature points, according to this can be deleted from those unnecessary features, but this has limitations, to the image according to the size of the fusion, to modify parameters, self-adaptability is not good.

The principle is very simple, the following direct, the shape of my Jiaotong university.

Panorama picture Stitching