Recently, LIDAR (Light Detection and Ranging) technique is in widespread use for obtaining a large number of points with three-dimensional coordinates. Detection and extraction of the objects from LIDAR data is needed for practical use. The LIDAR data, however, do not explicitly contain any geometric information, and features of objects implicitly exist in point clouds. The features, such as planes, lines and corners, can be only indirectly extracted by segmentation algorithms. In this article, we present a two-step algorithm based on the tensor voting framework for the extraction of planar features. First, we infer the normal vector at each LIDAR point by the plate tensors derived from geometric relations among the LIDAR points. In the second step, these normal vectors are classified by the density cluster method based on the directions of normal vectors. The density clusters can be divided into sub-clusters if the coordinates of points are introduced. The normal vectors at points in the same sub-clusters possess the same planar feature. The results of our experiments show that our algorithm is very effective in the automatic extraction of planar features from both airborne and terrestrial LIDAR data.