Self-supervised Lunar Surface Image Feature Point Detection and Matching

Traditional visual odometry is difficult to extract features in the sparse texture environment of the moon surface. In this paper, a self-supervised learning convolutional neural network is proposed for feature extraction and a Sinkhorn-based matching approach for feature matching. We use homography and Gaussian blur to construct the self-supervised learning dataset to solve the difficulties of data labeling. Feature matching based on the score of interest points further improves the matching accuracy. Compared to other traditional feature points such as ORB, SHIF, our algorithm, trained on pybullet-built lunar surface simulation images, can achieve accurate matching with better robustness. Compared with other learning algorithms like SuperPoint, our algorithm has an advantage in speed and has less error in homography.