Aiming at the problem that the 128-dimensional feature descriptors in scale-invariant feature transform（ SIFT） algorithm results in long time-consuming and poor real-time performance

an nonlinear improved SIFT algorithm based on local linear embedding（ LLE） is proposed

it is possible to preserve nonlinear information in the original data space

shorten the running time of the algorithm

and improve the matching precision. Secondly

aiming at the problem of large calculation quantity of Euclidean distance in the registration process

use the Manhattan distance to calculate the similarity between reference map and the image to be matched to further reduce the time spent by the algorithm. The experimental results show that the algorithm has high matching speed

while has high matching rate.