Toward Faster and Accurate Detection of Craters

Abstract

Impact craters are the most frequent geological features that a spacecraft may encounter while landing or navigating on planetary surfaces like those of Mars. A spacecraft’s Terrain-Relative Navigation (TRN) can be improved by automated extraction of these craters in real time. It is crucial to accurately pinpoint craters in the event of a soft landing, especially the smaller ones. This calls for an automated pipeline, which provides a fast and accurate detection of craters. The present research work makes use of the improved You Only Look Once (YOLO) version 8 for detecting craters from a Martian image dataset. The dataset is split into train, validation, and test sets, and the training set is also augmented for inducing variance into the model. The efficacy of the activation functions is tested by replacing the original SiLU from the YOLO backbone with ReLU, Mish, and SoftPlus activations. The unaltered backbone in itself manages to achieve very high accuracy as compared to the state-of-the-art (SOTA) techniques that have previously been used for crater detection. Mish activation shows the highest test ${F}1$ -score of 0.915, which is better than the originally used SiLU. The Mish-YOLO v8 manages to extract an average of 25 craters from single images in just 8.5 ms, which makes it the fastest existing crater detection pipeline having an ${F}1$ -score higher than 0.9, thereby making the proposed approach an excellent tool for automated crater detection from images.