To address the challenges of obstacle detection for lunar surface probes, this paper presents an innovative autonomous unmanned lunar surface obstacle detection and reconstruction system based on improved YOLOv9. By incorporating the attention mechanism SKNet and adopting the lightweight convolutional network structure StarNet into the YOLOv9 algorithm, the system enhances precision, recall, and mAP by 2.5%, 2.2%, and 1% respectively, while reducing parameters by 19.34% and GFLOPs by 21.53%. Additionally, the system integrates binocular vision technology with the SIFT feature extraction and matching algorithm, utilizing obstacle images and prediction box information generated by the target detection algorithm to achieve 3D reconstruction of rocks and craters on the lunar surface. The system not only improves the accuracy and real-time of obstacle detection and realizes the autonomous unmanned environment sensing of the rover on the lunar surface, but also provides certain technical support for the lunar exploration mission.