Abstract:In the context of long-distance exploration missions of lunar rovers facing harsh environments and threats posed by rock obstacles, this paper addresses the real-time obstacle avoidance problem. It involves the collection of ground-like lunar surface environmental data, manual screening, and labeling of image data using an embedded Segment Anything Model (SAM) tagging tool to construct a dataset of lunar-like surface rock image obstacles. Building upon this dataset, a method of fusing information from laser radar and monocular images for target perception is proposed. The YOLO V5 object detection network with embedded attention mechanisms operates on this dataset to generate a model for identifying lunar surface rock obstacles. Simultaneously, the laser radar collects environmental point cloud data and, following slicing, noise filtering, segmentation, and using the Random Sample Consensus (RANSAC) algorithm, produces point cloud obstacle data within the Point Cloud Library (PCL). By fusing the generated image obstacle data with the point cloud obstacle data, accurate obstacle information is determined, and distance data is extracted from the point cloud data. This approach enhances the accuracy and reliability of the lunar rover's environmental perception system, laying a robust foundation for real-time obstacle avoidance.