Empirical Analysis of Feature Points Extraction Techniques for Space Applications

Abstract

Recently, space research advancements have widened the scope of many vision-based techniques. Computer vision techniques with manifold objectives require that valuable features are extracted from input data. This paper attempts to analyze known feature extraction techniques empirically; Scale Invariant Feature Transform (SIFT), Speeded up robust features (SURF), Oriented fast and Rotated Brief (ORB), and Convolutional Neural Network (CNN). A methodology for autonomously extracting features using CNN is analyzed in more detail. The autonomous process demonstrates the use of convolutional neural networks for feature extraction. Those techniques are studied and evaluated empirically on lunar satellite images. For analysis, a dataset containing different affine transformations of a video frame is generated from a sample lunar descent video. The nearest neighbor algorithm is then applied for feature matching. For an unbiased evaluation, a similar process of feature matching is repeated for all the models. Well-known metrics like repeatability and matching scores are employed to validate the studied techniques. The results show that the CNN features showed much better computational efficiency and stable performance concerning matching accuracy for lunar images than other studied algorithms.

Soft Landing Parameter Measurements for Candidate Navigation Trajectories Using Deep Learning and AI-Enabled Planetary Descent

“…Few works describe existing vision-based deep learning navigation techniques for orbital and landing missions. e analysis in [22,23] discusses stateof-the-art feature extraction methods for deep learning applications. Depending on data availability, AI-based autonomous landing systems use deep learning approaches or deep reinforcement learning [24].…”