Topological-Projection-Consistent Keypoint Extraction for PnP-Based Space-Target Pose Measurement Under Extreme Illumination

Accurate Perspective-n-Point (PnP) pose measurement for space targets benefits from consistency between the topological structure of extracted two-dimensional feature points and that of projected three-dimensional template points. Existing keypoint extraction methods mainly optimize individual point localization accuracy, while paying insufficient attention to the global topological consistency of the keypoint set under relative-position relations, structural dependencies, and rigid projection constraints. As a result, frontend keypoint errors can be further amplified by the backend PnP solver. To address this problem, this paper proposes a topological-projection-consistent keypoint extraction method for PnP-based space-target pose measurement under extreme illumination. First, a Residual-Gated Context-Chained Enhancement Module (RG-CCEM) is introduced to build global structural feature propagation in degraded images. During distillation, both two-dimensional pairwise structural constraints and three-dimensional weak-perspective template reprojection constraints are imposed so that the student network learns keypoint sets whose projected topology remains consistent with the rigid 3D template. A reliability–reprojection-residual jointly reweighted PnP method is then used to reduce the influence of topologically inconsistent correspondences on pose estimation. Experiments show that the proposed method improves the stability of keypoint topology and pose measurement accuracy under extreme illumination. Compared with the DETRPose baseline, the proposed method decreases the outlier-keypoint ratio from 21.0% to 9.2%, and reduces rotation, translation, and camera-center RMSE by approximately 55.1%, 39.8%, and 56.3%, respectively, on our laboratory semi-physical dataset. Limitations of this method are further discussed.