Mapping is a very critical issue for enabling autonomous driving. This paper proposes a robust approach to generate high definition maps based on LIDAR point clouds and post-processed localization measurements. Many problems are addressed including quality, saving size, global labeling and processing time. High quality is guaranteed by accumulating and killing the sparsity of the point clouds in a very efficient way. The storing size is decreased using sub-image sampling of the entire map. The global labeling is achieved by continuously considering the top-left corner of the map images as a reference regardless to the driven distance and the vehicle orientation. The processing time is discussed in terms of using the generated maps in autonomous driving. Moreover, the paper highlights a method to increase the density of online LIDAR frames to be compatible with the intensity level of the generated maps. The proposed method was used since 2015 to generate maps of different areas and courses in Japan and USA with very high stability and accuracy.