Pose estimation for non-cooperative spacecraft involves determining the spatial position and attitude of spacecraft that lack active cooperation devices, such as defunct satellites or space debris. This technology is critical for advanced space applications, including autonomous rendezvous and docking, on-orbit servicing, and orbital debris removal. For non-cooperative spacecraft with unknown geometric structures, a prominent research approach involves leveraging simultaneous localization and mapping (SLAM) to reconstruct the three-dimensional (3D) structure using model-based methods for pose estimation. Current methods often rely on feature-based techniques to establish the pose constraints, followed by pose graph optimization to minimize cumulative estimation errors. However, the time-consuming nature of feature extraction poses challenges for real-time applications, and existing information matrices may inadequately represent pose estimation uncertainties. Sensors for pose estimation can be categorized as passive or active. Passive sensors are low-cost and high-frame-rate but can be affected by lighting variations. In contrast, active sensors like lidar directly acquire 3D point clouds, offering high accuracy and being less susceptible to lighting and scale variations. Therefore, we utilize lidar as the sensor for pose estimation. To enhance real-time performance and model reconstruction accuracy, we propose a non-feature-based 3D reconstruction and pose estimation method (NFRPE-3D) using lidar point cloud data.

First, we apply the iterative closest point (ICP) algorithm to execute a keyframe registration technique to obtain the relative pose, which is then recursively used to estimate the pose of the current frame. However, this process can introduce cumulative errors. To mitigate these errors, we update the pose graph based on attitude relationships between keyframes, establishing loop constraints. Pose graph optimization is performed using the g²o framework. Notably, the loop constraints in the pose graph are established solely through attitude relationships, which eliminate the need for complex feature extraction and reduce computational overhead. To address the limitation of existing methods where the information matrix does not accurately reflect pose estimation uncertainty, we propose constructing the information matrix for graph optimization by minimizing the sum of squared distances between corresponding keyframe points, thus enhancing pose graph optimization accuracy. The optimized pose graph results are then used to reconstruct the target’s point cloud model. After model reconstruction, subsequent pose estimations are performed using a model registration strategy.

To validate our method, we conduct a semi-physical simulation experiment using a 1∶1 satellite model under simulated space lighting conditions. We first evaluate the attitude and position estimation accuracy of various methods (Figs. 9?12, Table 2). Our methods achieve mean absolute errors of 2.34°, 1.67°, and 1.71° for the three-axis attitude, and 0.033, 0.007, and 0.025 m for the three-axis position, significantly outperforming other methods. Compared to existing feature-based methods, our method improves three-axis attitude and position accuracy by over 40%. A comparison of point cloud models before and after pose graph optimization (Fig. 13) shows that the proposed method effectively reduces cumulative errors and enhances model reconstruction accuracy. The reconstructed point cloud model (Fig. 14) delineates the overall structure, demonstrating the effectiveness of our model reconstruction step. Finally, the computational times of the methods are presented (Fig. 15). Overall, the maximum real-time computation time of our method does not exceed 0.2 s, and except for the model reconstruction step, the computation time remains below 0.1 s. The average computation time per frame is 0.040 s, demonstrating excellent real-time performance. Compared to existing feature-based methods, our approach increases the average computing speed by 95.8%.

In this paper, we propose a method named NFRPE-3D for point cloud model reconstruction and pose estimation of non-cooperative spacecraft using lidar point cloud data. The method establishes pose graph constraints based solely on attitude relationships between keyframes, eliminating the need for feature extraction and matching, which significantly reduces computational complexity. Furthermore, by minimizing the sum of squared distances between corresponding points in keyframes, we construct the information matrix for pose graph optimization, thus improving pose estimation accuracy. Experimental results demonstrate that NFRPE-3D effectively enables spacecraft model reconstruction and pose estimation. In the absence of target model information, pose estimation is dependent on keyframe registration, which can result in significant fluctuations due to cumulative errors. However, pose graph optimization effectively mitigates these errors, improving accuracy and stabilizing pose estimation. After model reconstruction, the model registration strategy further stabilizes pose estimation. Compared to existing feature-based methods, our approach improves three-axis attitude and position accuracy by over 40%, while also increasing the average computation speed by 95.8%.