Automatic recognition of diverse 3-D objects and analysis of large urban scenes using ground and aerial LIDAR sensors

We describe a learning-based 3D object recognition pipeline developed under the DARPA URGENT program for analyzing a large LIDAR dataset collected by both airborne and ground platforms for an extended urban area. Our approach utilizes a novel strip-based cueing approach that incorporates the properties and context of urban objects. Strip-based cueing segments potential objects and assigns them to appropriate classification stages. Our learning-based recognition pipeline successfully recognized 17 3D object classes in LIDAR data collected in and over Ottawa, Canada with high efficiency and average accuracy of 70%.