Moment-Net: A New Moment-Matching Neural Network for Pose Estimation

Abstract

Determining a spacecraft's pose is an important task for autonomous navigation. Conventional AI systems rely on peak detection and look only for the most intense pixel. Consequently, the tracking image frequently suffers from peak flattening in the presence of blur or other disturbances. This paper presents Moment-Net, a new neural network architecture based on L-moment matching that is designed to address this issue. Moment-Net treats the entire spatial distribution as a continuous geometric shape, rather than detecting keypoints based on a single maximum value. Moment-Net reduces complex 2D image data into independent 1D marginal probability distributions across horizontal and vertical channels, thereby lowering computational complexity for resource-constrained space-based hardware. L-moments up to the eighth order are extracted from these 1D distributions to capture fine spatial details. Furthermore, to dynamically assign weights to these L-moments, a global context-driven Multi-Layer Perceptron (MLP) is implemented, followed by a power-law decay filter designed to penalize higher-order L-moments that are usually prone to degradation. The network is trained under nominal blur-free conditions and tested across four different scenarios, ranging from unseen blur-free images to situations involving rapid rotation and significant image blur. The results demonstrate much superior localization accuracy compared to standard HRNet.