Abstract: In distant retrograde orbit (DRO) optical autonomous navigation, traditional image-processing-based centroid extraction methods often suffer from severe performance degradation or even failure under extreme illumination and phase conditions, such as crescent and gibbous phases. These operating scenarios are characterized by significant target occlusion and non-uniform illumination, under which edge-based detection and fitting approaches become highly sensitive to noise and prone to convergence toward local optima, thereby constituting a major limitation for high-accuracy angle measurements. To address this challenge, this paper proposes a dual keypoint regression method based on deep geometric constraints, in which the conventional centroid extraction problem is reformulated as a feature learning task that jointly incorporates physical and geometric priors. The proposed method adopts YOLOv8-pose as the keypoint regression framework and introduces the physical centroid and the geometric center as jointly regressed targets. By exploiting the geometric constraints implied by the observable illuminated arc, the prediction space of the centroid in unobservable regions is effectively restricted, leading to enhanced robustness under extreme phase conditions. Validation experiments conducted on an on-orbit observation dataset generated by a high-fidelity virtual simulation engine demonstrate that, compared with the traditional ellipse-based direct method, the proposed approach reduces the angular measurement error decreases from 0.116° to 0.013°, corresponding to an order-of-magnitude improvement in angular accuracy. These results confirm the effectiveness and feasibility of the proposed method in complex deep-space imaging environments, providing a viable solution for high-reliability optical autonomous navigation.