Monocular 6-DoF Pose Estimation of Spacecrafts Utilizing Self-Iterative Optimization and Motion Consistency

Abstract

Monocular 6-DoF pose estimation is crucial for spacecrafts to achieve precise navigation and positioning, and it has gained increasing attentions in recent years. However, spaceborne imaging quality is heavily influenced by specific factors such as varying illumination conditions, low signal-to-noise ratio and high contrast. In addition, the lack of sufficient labelled space data hampers the performance of deep learning-based pose estimation methods. To overcome these challenges, we propose a novel monocular 6-DoF pose estimation method for spacecrafts utilizing self-iterative optimization and motion consistency. Firstly, we reconstruct an initial 3D spacecraft model using manually annotated 2D keypoints from several images, which can generate the labels of 2D keypoints, heatmaps, and bounding boxes for the entire training set. Subsequently, we train a Multi-task Key-point Prediction Network (MKPNet) model using these label information, and through an iterative optimization process, refine both the 3D model and the performance of MKPNet in predicting 2D keypoints. Additionally, we incorporate temporal information and motion consistency from sequential images to smooth the pseudolabels of poses predicted by MKPNet during testing. This smoothing process guides the self-training process of the network model, leading to improved generalization and pose estimation accuracy. In the SPARK 2024 Challenge, our method achieves competitive results compared to the state-of-the-art methods and outperforms the baseline regression approaches by a significant margin.