Domain Adaptation and Multi-view Attention for Learnable Landmark Tracking with Sparse Data

TL;DR

This paper introduces a lightweight, real-time landmark tracking system for space missions that combines domain adaptation and multi-view attention to improve detection and description of celestial terrain features with limited data.

Contribution

It presents novel domain adaptation and attention-based neural network methods tailored for efficient, in-situ landmark tracking on spacecraft hardware.

Findings

Outperforms existing state-of-the-art landmark tracking methods

Enables robust detection with limited and cheaply acquired training data

Achieves real-time performance on current spacecraft processors

Abstract

The detection and tracking of celestial surface terrain features are crucial for autonomous spaceflight applications, including Terrain Relative Navigation (TRN), Entry, Descent, and Landing (EDL), hazard analysis, and scientific data collection. Traditional photoclinometry-based pipelines often rely on extensive a priori imaging and offline processing, constrained by the computational limitations of radiation-hardened systems. While historically effective, these approaches typically increase mission costs and duration, operate at low processing rates, and have limited generalization. Recently, learning-based computer vision has gained popularity to enhance spacecraft autonomy and overcome these limitations. While promising, emerging techniques frequently impose computational demands exceeding the capabilities of typical spacecraft hardware for real-time operation and are further challenged by the scarcity of labeled training data for diverse extraterrestrial environments. In this work, we present novel formulations for in-situ landmark tracking via detection and description. We utilize lightweight, computationally efficient neural network architectures designed for real-time execution on current-generation spacecraft flight processors. For landmark detection, we propose improved domain adaptation methods that enable the identification of celestial terrain features with distinct, cheaply acquired training data. Concurrently, for landmark description, we introduce a novel attention alignment formulation that learns robust feature representations that maintain correspondence despite significant landmark viewpoint variations. Together, these contributions form a unified system for landmark tracking that demonstrates superior performance compared to existing state-of-the-art techniques.