Mars topographic data are crucial for quantitative surface characterization, exploration missions, and studies of Martian surface processes. Photogrammetric processing of Mars orbital imagery is a majormethod for generating three-dimensional (3D) terrain models, such as digital elevation models (DEMs), with bundle adjustment (BA) serving as the key step for mitigating inconsistencies in overlapping regions of different orbital images and further improving the spatial accuracy of the resulting DEMs. However, BA performance is often limited by the texture-less Martian surface and the lack of ground control points. To address this issue, this paper proposes a BA method assisted by robust crater graph features. Since impact craters are widely distributed on Mars, they can serve as valuable semantic priors for accurate topographic mapping. The method first uses deep learning to extract crater structures and constructs crater graphs based on the minimum spanning tree rule. It then searches for corresponding crater graphs across different images to identify crater correspondences and robust tie points. Finally, angular relationships among adjacent craters are introduced into BA observation equations to enhance adjustment performance and reduce geometric inconsistencies. serve as valuable semantic priors for accurate topographic mapping. Experiments conducted over the McLaughlin Crater area using CTX stereo images, with HRSC-derived DTMs and orthoimages as reference data, demonstrate that the proposed method effectively improves the precision and stability of BA and supports high-accuracy 3D mapping of the Martian surface.