Near-Real-Time Detection of Craters: A YOLO v5 Based Approach

Abstract

In planetary surfaces like that of Mars, impact craters are the most commonly found geological structures that a spacecraft may encounter while landing or navigation. Automated extraction of these craters in real-time can improve terrain relative navigation of a spacecraft. It can improve precision of a spacecraft's position estimate by providing supplementary measurements to correct for drift in the inertial navigation system. It is also very important to precisely locate craters, especially smaller ones during the event of a soft landing. This calls for an automated system that can detect craters with high accuracy in near real time. To tackle this problem we propose to use a YOLO based detection because YOLO tackles detection as a single regression problem, straight from image pixels to bounding box coordinates and class probabilities. It is a single Convolutional Neural Network (CNN) simultaneously predicting bounding boxes and respective class probabilities for each bounding box. The YOLO network is trained on a crater dataset created from Mars Reconnaissance Obiter's CTX and tested on both CTX and Mars Express mission's HRSC images. The fully optimized CNN showed promising detection results, achieving state-of-the-art detection precision, with excellent recall at an astounding speed of 64.8 ms for 115 craters. © 2023 IEEE.