The optical images from lunar rover could be used to reconstruct the lunar terrain spatial configuration. The products include 3D point cloud and digital orthophoto map, etc. While the images matching algorithm accuracy is closely related to the reconstruction’s accuracy and reliability, an improved dynamic programming matching method is required. Considering bidirectional constraints of row and column simultaneously and adding the dynamic seed point strategy, this algorithm can get global optimal solution with the advantages of high efficiency and stability. Compared with the result from others methods and the ground laser data, experimental results demonstrate that the method has a high accuracy and capability to meet the demand of lunar rover positioning and navigation exploration project.