A Sequential Approach for On-board Rock Detection from Lunar Images

Rock segmentation in lunar images is a crucial computer vision task for visual navigation of planetary rovers. Even though numerous approaches have been already proposed to address this task, many solutions are underpinned with computationally-intensive deep learning models, which makes them unsuitable for on-board processing. In the study reported here, we address this important problem and we propose a sequential pipeline, which combines a U-Net-based network for rock segmentation with a YOLO model for final object detection. We demonstrate that putting two lightweight models together improves the detection performance, making it close to that obtained with full-sized architectures. Even though this is an initial study, the obtained results indicate that this direction is promising and it is worthy of further investigation.