Abstract: Aiming at the problem that the traditional LiDAR detection model is prone to misjudging the drivable area when there are obstacles, railroad tracks, curves and other interferences in the unstructured road without obvious edge features, a high-precision detection method for the drivable area of complex unstructured roads is proposed. Firstly, an optimized incremental random sampling consistency algorithm is proposed to extract the ground surface for the non-obvious elevation difference between the ground surface and the non-ground surface point cloud; then, the road direction is determined based on the beam model, and the point cloud feature matrix is constructed by integrating the point distance, the elevation difference of the cloud, and the direction vector of the cloud; furthermore, the point cloud feature matrix and the position of the LiDAR scanning line are integrated to select the left and right boundary features of the road; Finally, a Gaussian process regression method is used to predict the boundary points and solve the problem of road edge occlusion. The experimental results on KITTI data and self-constructed data show that the proposed method is applicable to both structured and unstructured roads, especially in the unstructured road scenarios, it is significantly better than other classical methods, with an accuracy of 85%; in several different road morphology scenarios, the method has both better accuracy and real-time performance.