Construction and Analysis of a Simulated Lunar Surface Image Dataset for Lunar Rock Detection

Lunar rovers are effective long-range scouting tools for future lunar missions. However, unknown complex environments and randomly scattered rocks could impede rover navigation and safety. Accurate obstacle detection is essential for identifying hazards and intelligent path planning. Research progress is partly hindered by limited public lunar image dataset. To facilitate rock detection for lunar rovers, we construct a lunar-like surface image dataset using selective field image collection and annotation with a Segment Anything Model (SAM) assisted labeling tool. The dataset contains 659 images, split into training, validation and testing subsets. The experimental results demonstrated that the YOLOV5 and YOLOV5 object detection network models with attention mechanisms achieved the highest Precision of 0.878 and mAP of 0.783 when tested on this lunar surface dataset. The dataset provides a testbed to develop and evaluate rock perception algorithms, advancing lunar rover autonomy and hazard avoidance capabilities.