In this study, we propose a ground-segmentation algorithm based on a two-stage coarse-fine processing approach to address the limitations of traditional LiDAR ground-segmentation algorithms, such as poor real-time performance and threshold dependence for complex slope roads. First, the point cloud was divided into fan-shaped regions, with the local slope threshold of each area being adaptively determined to complete the first stage of rough segmentation. Subsequently, the point cloud was projected onto the RGB image, and "unknown classification points" were obtained using an image expansion algorithm. Finally, effective points were screened out to perform convolution filtering on the "unknown classification points" to achieve fine segmentation and to determine the distance threshold by dividing multiple regions. The results demonstrate that the proposed algorithm achieves a ground-point segmentation accuracy exceeding 96% for both flat roads and complex slope roads. In addition, the recall rates consistently maintain a high level of over 95%. Moreover, the local slope threshold can be adjusted to achieve excellent robustness, and the average processing time is 16.57 ms, which satisfies the real-time requirements of unmanned vehicles.