In driverless technology, LiDAR is used to scan and detect the surrounding environment to achieve 3D imaging, target recognition, obstacle avoidance, map construction, and autonomous navigation. Among them, surface feature extraction based on laser scanning point cloud is conducive to the accurate 3D reconstruction of the surface shape of objects, which is a necessary prerequisite and important means for 3D object recognition, obstacle avoidance, and autonomous navigation. Therefore, to effectively extract the surface features of 3D objects from laser point clouds, a combined optimization algorithm for surface feature extraction is proposed in this paper, that is, in the sequence of point cloud processing, a variety of measures are integrated to achieve the optimal processing, to improve the reliability and accuracy of each link processing algorithm. The specific optimization measures are as follows. First, the RANSAC (random sampling consistency) sampling strategy is used to optimize the point neighborhood，Secondly, the Harris-3D algorithm is used to extract key points from point cloud data, and combined with the region growth method based on the angle of the normal vector and Euclidean distance double threshold, the point cloud is segmented by clustering. Finally, feature extraction of the 3D object surface is carried out on the point cloud surface after clustering and segmentation, and feature identification of the 3D object surface form is realized. Through experiments on the extraction and reconstruction of regular surfaces in point clouds, the results show that the proposed integration optimization algorithm can effectively improve the surface accuracy and efficiency of feature extraction of the regular surfaces of the 3D object from point clouds, for the plane, cylinder, cone, the extraction of quadric surface reconstruction error is less than 0.075 mm, and for the spherical surface, the reconstruction error is less than 2 mm. In addition, the experimental verification of the real unmanned laser scanning scene with a large amount of point cloud data is also carried out, which shows that the algorithm also has good surface feature extraction effectiveness, and can effectively realize the recognition and reconstruction of 3D object surface shape.