Aiming at the problems of the existing obstacle detection methods in complex road scenes, such as inaccurate ground segmentation, large amount of calculation and difficult target clustering at different distances, a roadside lidar based obstacle detection method was proposed. In the aspect of ground plane segmentation, the improved fan-shaped grid map based on cylindrical coordinate system and the lowest point representation method are proposed to optimize the selection of seed points. The ground fitting and segmentation are realized by using multi-plane model and random sample consensus algorithm ( RANSAC ). In the aspect of obstacle clustering, KDTree is constructed to accelerate the clustering process, and the Euclidean clustering algorithm is improved by dividing the region and threshold adaptive. The experimental results show that the segmentation accuracy of this method for ground points in four typical road scenes is more than 86%, and the clustering accuracy of obstacle targets at different distances is significantly improved.