To achieve the accurate automatic extraction of spherical targets from 3D laser point clouds in complicated scenes, a method was proposed for the automatic and accurate extraction of spherical targets based on SHOT features. This method designed the processes of rough extraction and refined extraction. In the rough extraction process, SHOT feature descriptor was utilized to extract all spherical target point clouds in the scene; secondly, Euclidean clustering was used to segment spherical target point clouds, and rough spherical target parameters was calculated using the least square. The refined extraction process was based on the iterative least squares method and normal filtering to eliminate the aspherical points, and obtain the spherical target point cloud and accurate spherical target parameters. An experimental scene with 4 spherical targets was designed. The German Z+F Image 5016 scanner was employed to collect the scene data, and the spherical target point cloud and spherical target parameters in the experimental scene were automatically extracted. The results show that in the range of 10 meters, the error of the radius of the spherical target automatically extracted by this method was 0.25–0.33 mm, which was 0.02–0.06 mm less than that of the manually extracted spherical target point cloud, and 0.01–0.09 mm less than that based on the differential method. The proposed method can achieve high positioning accuracy for spherical targets and robustly eliminate noise in the scene point cloud, and can complete the automatic extraction of spherical targets with millions of point clouds within 30 seconds.