Abstract: To ensure high precision, reliability and efficiency in spacecraft pipe system assembly, accurate target detection and pose estimation are essential. However, the untextured and complex geometry, as well as susceptibility to light and occlusion, leads to low accuracy of traditional texture- and geometry-based methods. To solve these challenges, a novel pipe pose estimation method based on adaptive convolution and latent representation is proposed. Firstly, diverse and realistic synthetic datasets are generated using domain random, mitigating challenges in large-scale data acquisition and annotation. Then, to address the problem of low segmentation accuracy due to the complex geometry of pipes, an adaptive convolution-based pipe instance segmentation network is proposed, dynamically adjusting to pipe complex geometry to better capture key features and spatial relationships. Structured pruning further optimizes network efficiency and segmentation accuracy. Finally, to address the problem that traditional positional features are limited in expression and susceptible to environmental changes, a novel latent pose representation is designed, which is only sensitive to pose transformation but robust to other factors. Supervised and self-supervised learning are integrated for initial pose estimation, followed by an edge feature-based optimization algorithm to enhance accuracy. Experimental results show that compared with the existing pose estimation methods, our proposed methods achieve more accurate pipe segmentation and pose estimation, with a segmentation accuracy of 99.2%, a speed of 32.8 FPS, and pose accuracies of 2.445 mm and 1.074°, which meets the requirements of target detection and pose estimation in automated pipe assembly.

Key words: pose estimation, instance segmentation, domain randomization, adaptive convolution, latent pose representation