Abstract: Aiming at the demand for safe obstacle avoidance in the autonomous navigation of Mars rover in complex terrain and the double constraints of computational resources and energy supply of the onboard platform, this paper constructs the YOLOv8-LMD lightweight detection model, aiming at realizing the requirements of high precision and lightweight characteristics of the rock detection algorithm on the surface of Mars. First, the lightweight backbone network is reconstructed based on the HGNetv2 architecture to realize the preliminary compression of model parameters. Secondly, a multi-scale feature fusion network structure is designed, and the neck network is reconstructed by integrating Slim-neck and ASF-YOLO to effectively improve the feature characterization of rock targets at different scales. In addition, a lightweight detection head is designed by using the convolutional sharing strategy, which reduces the computational complexity and enhances the classification and localization accuracy at the same time. Finally, a pruning algorithm is used to prune the model parameter redundancy to further compress the model, and the knowledge distillation technique is used to achieve the compensation and optimization of the accuracy. Through experiments, it is found that compared with YOLOv8n, YOLOv8-LMD accuracy is improved by 1.7%, the computational amount is reduced by 68%, the parameter amount is reduced by 77%, and the model size is reduced by 75%. Therefore, it can be considered that the model in this paper is more suitable to be applied in the Mars surface rock detection task.