Because the feature points method is easier to deal with the transformation between images, such as rotation, affine, perspective and so on, it is often used. The feature points include the corners of the images and the points of interest that show some singularity relative to their fields. Harris et al. proposed a corner detection algorithm, which is a well-known corner detection algorithm with rigid transformation invariance and affine transformation invariance to a certain degree, but the algorithm does not have a zoom transformation invariance. For such shortcoming, Lowe proposed an SIFT feature points with zoom invariance.
As shown in Figure 5, the image stitching requires finding valid feature matching points in the image sequence. The search for feature points of the image directly affect the accuracy and efficiency of image stitching. For the image sequence, if the number of feature points ≥ 4, it is easy to automatically calibrate the image matching points; if the feature points are few, image stitching often can not achieve more ideal results.