Dense Structure Inference for Object Classification in Aerial LIDAR Dataset Eunyoung Kim and G´ erard Medioni Institute for Robotics and Intelligent Systems University of Southern California Los Angeles, CA, USA {kimeunyo | medioni}@usc.edu Abstract—We present a framework to classify small free- form objects in 3D aerial scans of a large urban area. The system first identifies large structures such as the ground surface and roofs of buildings densely built in the scene, by fitting planar patches and grouping adjacent patches similar in pose together. Then, it segments initial object candidates which represent the visible surface of an object using the identified structures. To deal with sparse density in points representing each candidate, we also propose a novel method to infer a dense 3D structure from the given sparse and noisy points without any meshes and iterations. To label object candidates, we build a tree-structure database of object classes, which captures latent patterns in shape of 3D objects in a hierarchical manner. We demonstrate our system on the aerial LIDAR dataset acquired from a few square kilometers of Ottawa. Keywords-Object classification; Range image; LIDAR; Den- sification; I. Introduction As recent advances in light detection and ranging(LIDAR) technology allow the ability to collect 3D data over vast urban areas with excellent resolution and accuracy, the auto- matic recognition of 3D small objects in the scene becomes important for various applications such as environment monitoring and autonomous robotic navigation. Our goal is thus to develop a system that automatically categorizes small free-form 3D objects(e.g. cars) in aerial range images, whereas the majority of existing works on large-scale range image processing have focused on identifying terrain and buildings or linear structures such as poles. The first image of Fig. 1 shows an example of typical urban region containing buildings, houses, vegetation and small objects. The region covers 2, 196 × 2, 997m 2 and produces the number of 3D points(2.5GB in binary). Specif- ically, there are up to a few hundreds of thousands of 3D points in 50 × 50m 2 region. Labeling small objects in an aerial LIDAR data is challenging because of irregularity and sparsity in point clouds representing the visible surfaces of 3D ob- jects(approximately, 1 point in 20 × 20cm 2 region). Fig.2(a) shows the original point cloud from a car, which has sparse 3D points and a big hole caused by a front window. This may result in misinterpreting the shape of an object and, 2,196 m 2,997 m Wide area view Aerial LIDAR data Ground(white) & roof identification Candidate object analysis (Dense structure inference) Object classification (red box: identified cars) Initial object segmentation Figure 1. System overview consequently, poor classification performance. Therefore, our system also includes dense structure inference from given sparse 3D points. Fig. 1 shows a flow chart of our system. Given an aerial LIDAR data, the system starts by identifying the terrain and the roof surfaces of the buildings(houses) using planar primitives, and delineates the initial object candidates using spatial contexts between the large structures and small objects. Then, for every candidate, we infer uniformly-sampled dense 3D points smoothly continuous with the existing surface from the given sparse point cloud. Finally, the system labels each resultant point cloud which represents the visible surface of a 3D object. Our contribution is two-fold: 1) We develop a generic system that recognizes free-form 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.747 3041 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.747 3053 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.747 3049 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.747 3049 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.747 3049