Since multi-cameras images involve much differences in spatial characteristics and spectral characteristics, so it is full of difficulties in the image registration. According to the different characteristics of the multi-cameras images, this paper proposed a new algorithm of sub-pixel image registration based on Harris corner and Scale Invariant Features Transform (SIFT) descriptor. The algorithm consists of three procedures: feature detection, pixel-level registration and sub-pixel-level registration. Firstly, the Harris algorithm was selected to extract the feature corners and determine the main direction of the Harris corners. Secondly, the SIFT descriptor was chose to describe the key points. Then, feature points acquired on matching by the two-way nearest neighbor algorithm. Finally, in the sub-pixel-level registration process, we carry out interpolation in the neighborhood of the pixel-level matching points. Then the pixel-level registration is taken once again. The experimental results show that, the proposed algorithm is accurate, efficient, and retains the rotational invariance of the SIFT descriptor. What’s more, processing speed is significantly increased.