Reprinted from: http://blog.csdn.net/yuyin86/article/details/6690423
Panoramic video is a new method of creating virtual environments using 360-degree panoramic images. Panoramic images are created by stitching images with overlapping borders taken by ordinary cameras. The similarity of the corresponding pixels in the overlapping part of the image can be used to make the obtained image seamless and smooth by adopting an effective stitching algorithm.
Posts from research forums Walkfarer and SCQ:
http://bbs.matwav.com/post/view?bid=6&id=371051&sty=3&age=0&tpg=1&ppg=1#371051
Image stitching is an important branch in computer vision, which is to seamlessly stitch two or more partially overlapping images to obtain images with higher resolution or wide viewing angle. The principle of the current shaping algorithm is roughly as follows:
1. Frequency domain: (not very clear)
Use the "phase correlation method" to complete the translation estimation of the two images (only accurate to the pixel level).
2. Spatial domain:
(1) Feature -based method: find feature points and match them.
a. Use HARRIS corner points, and then use grayscale correlation plus relaxation matching to find corresponding points. If the overlapping area of the two images is large and the perspective distortion is small, this method can be considered.
b. Determine a set of optimal feature matching through information such as the gradient direction of the feature points, use this set of data to give the estimated initial value of the transformation matrix between the two images, and then use the recursive algorithm to find the final accurate transformation relationship. Better results can be achieved when the rotation and zoom of the camera that takes the picture is not very large.
(2) and light-difference -based methods (methods are accurate but slow to converge). One of the most widely used stitching algorithms is the stitching of cylindrical and spherical images. After spherical and cylindrical transformations, the problem boils down to determining the amount of translation for each image.
The above is a summary of some articles I read, which is laughable and generous. I hope you will correct or add to it.
*********csq************
Let me add it.
1 Frequency domain: Generally, the phase correlation of fourier is used to estimate the frequency shift, rotation and scaling.
There is nothing to discuss about frequency shift. It is relatively simple. It is estimated that scaling and rotation are usually done by changing to the polar coordinate system. Countless papers discuss this issue. The previous post on "How to judge whether two images overlap", I gave a This paper, this year's ieee tran ip. not bad. Some people in the frequency domain do sub-pixel. I have also said this, the faroon of the university of centrl florida (can't remember the name), wrote an article like this, but it seems that it can only handle the frequency-shifted sub-pixel. I also emailed him to ask if he handles rotated sub-pixels. He didn't reply. ft.
2. Airspace:
1. The landlord missed the method based on intensity, but it is really not used much.
2. Feature method:
It seems that the most common is point-rgistration. However, there are still people who use the edge to fight with other features.
Now I will introduce the point hair.
Point-registration:
1．经典的harris point, 现在有不少修正版 ，因为harris 当时对参数的选择并没有给出很好的建议，所以参数选择比较烦，我以前就试过很多参数.
参考文献：A Mathematical Comparison of Point Detectors
Evaluation of Interest Point Detectors
找到特征点那就要去匹配拉。
一般是先初步估计一下，剔出差太多的匹配对。 可以用intensity的各种方法，但是这就摆出一个问题： 光照。 很麻烦，两幅图的光照差很多的话，有可能根本弄不出来匹配的点，但你从图上明显可以看到很多对都是匹配的。这个等会讨论。
2．然后进一步估计匹配，RANSAC用的最多，也有其他的方法， 像paper: MLESAC- A new robust estimator with application to estimating image geometry.
估计出匹配对，然后就要算那个乱七八糟的矩阵。 又是很多方法可以来做。一般是各种优化算法像 LM之类的。
3．算出矩阵，然后把一个变到另外一个的坐标系，就是融合的问题了。咋个无缝拼接是个问题。 Szeliski的方法用的多，paper我忘了，等会那篇review里面可以查到。
没有考虑的问题：
1． 假设是perfect 的各种关系，像频移，旋转，缩放，仿射，实际上并不那么简单，实际数码照的片子关系很复杂。
2． 光照。刚才说了，大问题。在估计匹配点，矩阵，甚至最后融合都会引入不少麻烦。
SIFT 方法好，可以解决一些问题，
经典的一篇综述
Image alignment and stitching- A tutorial
http://hi.baidu.com/simonyuee/blog/item/24961f6dfaa543fa431694bd.html