Outline of Presentation Automated Feature Extraction from Terrestrial and Airborne LIDAR Presented By: Stuart Blundell Overwatch Geospatial - VLS Ops Co-Author: David W. Opitz Overwatch Geospatial - VLS Ops GEOBIA 2008 7 August 2008 Introduction to Overwatch Geospatial Software Feature Analyst and LIDAR Analyst Software Automated Feature Extraction from LIDAR Challenges in Object Detection and Identification Research in Urban 3D Feature Extraction Classification Algorithms Meshing and Iso-Surfacing Approach Output of 3D Urban Models Future Capabilities and Conclusions Overwatch Geospatial Feature Analyst Software Machine learning approach for assisted to automated feature extraction Database Features RemoteView ELT 5500/3500 Feature Analyst Cost of geospatial database maintenance is 60-80% of system costs LIDAR Analyst Urban Analyst InterSCOPE Imagery and LIDAR are source data for geospatial features Feature Analyst provides an adaptive, integrated solution GIS Users Images & LIDAR Feature Analyst LIDAR Analyst Software Feature Type: Buildings Extraction Time: 00:02:56 Number of Buildings: 843 Editing Time: 00:01:10 (Convert to Point) Feature Analyst AFE Model Building Extraction from QuickBird multi-spectral satellite Imagery. Building features are collected as polygons then converted point vector features. Developed in 2004 as a plug-in for ArcGIS and ERDAS Imagine Challenge: Automate the extraction of Bare Earth surfaces, Buildings and Trees from LIDAR Motivation: Deliver GIS ready data layers to support terrain and 3D feature analysis Ongoing tool and algorithm development to meet the evolving requirements of our customers 1
LIDAR Data: Introduction Used to map topography, elevation of raised surface features Collected in a S or Z pattern in a 30 degree-wide swath Elevation is calculated by measuring time required for signal to travel to object and reflect back to sensor Each return contains at a minimum XYZ values LIDAR Data: Characteristics Raw LIDAR output is commonly known as a point cloud Returns (or reflections) occur from various interesting objects power lines buildings vegetation Ground Format is usually ASCII, and increasingly LAS Challenges: Feature Extraction Multi-Sensor Views of Objects Software development is always chasing after advances in sensor capability: Multi-sensor data collection RGB data with LIDAR points Higher spatial resolutions & very large datasets Feature extraction requires: Moving beyond visualization of points Object detection and identification Extraction of feature geometry, texture and descriptive attributes Spatial context and semantic knowledge for more complex features i.e. gas station Left Sensor View Center Sensor View Right Sensor View Registration of Multi-Sensor Views RGB and LIDAR Points Increase in number of points over the areas that are common to the three sensors Can lead to redundant information and bloat in point cloud size Even though the 3 sensors are located on the same platform the time of capture occurs is different Point Cloud of All Three Views 2
Object Detection & Identification Typology of 3-D Urban Models [Shiode 2001] X,Y, Z Values Only X,Y, Z, Intensity and Context Research Project Overview AFE Challenges in the Urban Environment Dynamic collection environment for vehicular based terrestrial LIDAR systems: Cluttered environment that impacts collection of data (occlusions) due to traffic, construction, and other factors Complex feature representation from street-level view of objects as a result of imaging angles and state of feature i.e. size, shape, context and condition of feature Very-large data volumes and lack of metadata standards for sensor path and other collection parameters Minimal collaboration between AFE software community and LIDAR hardware community Product Vision Urban 3D Modeling Toolkit: LIDAR Analyst: Terrestrial LIDAR AFE Airborne LIDAR AFE 3D Feature Geometry, 3D Feature Geometry, Textures and Attributes Textures and Attributes (Street View) (Map View) 2D and 3D Feature Material Identification Urban 3D AFE Geometry, Textures and Mapping and Attributes 3D Visualization Hyperspectral Toolkit: and Analytics Feature Analyst: Feature Analyst AFE from Imagery and Full Motion Video Urban Analyst Urban 3D Modeling Toolkit Current Capabilities TOC to manage layers Load point clouds (LAS) Visualize data in 3D Basic 3D Navigation tools Displays points, ground TIN, models Color points by elevation, intensity, classification Run classification algorithms and display results Classification Algorithms Current extraction algorithms and output format: Ground (points or mesh) Buildings (points, simple geometry, or mesh) Vehicles (points) Trees (points or models) Poles (points) Windows / Doors (points or holes in simple geometries) Urban 3D Modeling Toolkit 3
Tensor Voting Method Tensor Voting is a more robust and efficient method compared to a hierarchical approach Implements a spatial-partitioning data structure and uses a regional density clustering algorithm Slightly higher memory footprint than an Oct-Tree but allows for much faster processing Internal storage mechanism is similar to a voxel matrix with a variable Level Of Detail (LOD) and regional point statistics All processing is point based rather than voxel based to produce greater accuracies Visual Representation of a Tensor Ground High surface saliency Low point saliency Normal pointed up Wall High surface saliency Low point saliency Normal pointed sideways Tree High point saliency Benefits of the Tensor Approach Allows matching of features with greater variability than previous methods Robust to noise at large scales Adaptable to varying point density Supports the extraction of multiple features in a single pass Captures more information during extraction which can be translated into feature attributes Challenges of Using Terrestrial Data Points on a wall deviate a maximum of 30 centimeters 3D View Overhead View Profile View Resulting Meshes from Bad Data Useful Data Characteristics Accurate registration or collected from a single pass Building facades have a single thin plane of points Good point density Overhead View 4
Building Façade Model Generator Use extraction algorithms to segment building points from non-building points Triangulation is computed from building points Open Flight mesh is created from triangulation to maximize detail of the facades Meshing Example Point Cloud Mesh Smithsonian Point Cloud Resulting Open Flight Model Meshed facade Meshing and Iso-Surfacing Delaunay Triangulation converts points lying on a plane into a mesh of triangles Triangulation produces no overlapping triangles Creates Open Flight mesh for building facades and TINs for ground surface Iso-Surfacing algorithms can create a mesh around spherical objects like trees Can result in models with a high number of triangles Better to use simple geometries or a pre-built models for visualization efficiency Complete 3D Building Models Use registration algorithm to register terrestrial point cloud to aerial point cloud Required since point clouds do not exactly overlap due to sensor error Use LIDAR Analyst to extract building footprints and U3DMTK to extract building facades A matching algorithm will match facades to footprints Matched facades will be inserted into aerial models Missing facades will be extruded using attributes collected from aerial data Combined Terrestrial-to-Aerial LIDAR Point Clouds Complete 3D Building Model Example Aerial Photograph Building Footprint from Aerial LIDAR Close up view of combined point cloud Building Facades from Terrestrial Final Building Model 5
Automated 3D Urban Models Integrate terrestrial result with aerial results Determine representation for trees and vehicles Future Capabilities and Conclusions Feature Extraction from Airborne LIDAR Increased levels of automation Specialized tools for finishing 3D data to user specification i.e. (geometry, texture, attributes, behavior) Spatial awareness of feature level data Feature Extraction from Terrestrial LIDAR Data fusion from multi-sensor platforms to support object detection and extraction Co-registration of aerial and terrestrial LIDAR Appropriate 3D Data Structures and Models Contextual knowledge algorithms with user input to guide object identification i.e. dumpsters Questions and Open Discussion 6