Frank Stolle

The Ascender System

The Ascender system acquires, extends, and refines 3D geometric site models from calibrated aerial imagery. To acquire a new site model, an automated building detector is run on one image to hypothesize potential building rooftops. Supporting evidence is located in other images via epipolar line segment matching in constrained search regions. The precise 3D shape and location of each building is then determined by multiimage triangulation under geometric constraints of 3D orthogonality, parallelness, colinearity, and coplanarity of lines and surfaces. Projective mapping of image intensity information onto these polyhedral building models results in a realistic site model that can be rendered using virtual “fly-through” graphics. As new images of the site become available, model extension and refinement procedures are performed to add previously unseen buildings and to improve the geometric accuracy of the existing 3D building models. In this way, the system gradually accumulates evidence over time to make the site model more complete and more accurate. An extensive performance evaluation of component algorithms and the full system has been carried out. Two-dimensional building detection accuracy, as well as accuracy of the three-dimensional building reconstruction, are presented for a representative data set.

Error detection and DEM fusion using self-consistency

The ability to efficiently and robustly recover accurate 3D terrain models from sets of stereoscopic images is important to many civilian and military applications. Our long-term goal is to develop an automatic, multi-image 3D reconstruction algorithm that can be applied to these domains. To develop an effective and practical terrain modeling system, methods must be found for detecting unreliable elevations in digital elevation maps (DEMs), and for fusing several DEMs from multiple sources into an accurate and reliable result. This paper focuses on two key factors for generating robust 3D terrain models, (1) the ability to detect unreliable elevations estimates, and (2) to fuse the reliable elevations into a single optimal terrain model. The techniques discussed in this paper are based on the concept of using self-consistency to identify potentially unreliable points. We apply the self-consistency methodology to both the two-image and multi-image scenarios. We demonstrate that the recently developed concept of self-consistency can be effectively employed to determine the reliability of values in a DEM. Estimates with a reliability below an error threshold can be excluded from further processing. We test the effectiveness of the methodology, as well as the relationship between error rate and scene geometry by processing both real and photo-realistic simulations.

Site model acquisition and extension from aerial images

A system has been developed to acquire, extend and refine 3D geometric site models from aerial imagery. This system hypothesizes potential building roofs in an image, automatically locates supporting geometric evidence in other images, and determines the precise shape and position of the new buildings via multiimage triangulation. Model-to-image registration techniques are applied to align new, incoming images against the site model. Model extension and refinement procedures are then performed to add previously unseen buildings and to improve the geometric accuracy of the existing 3D building models.<<ETX>>

Cost-Effective Determination of Biomass from Aerial Images

This paper describes an ongoing collaborative research program between the Computer Science and the Forestry and Wildlife Management Departments at the University of Massachusetts to develop cost-effective methodologies for monitoring biomass and other environmental parameters over large areas. The data acquisition system consists of a differential GPS system, a 3-axis solid state inertial reference system, a small format (70mm) aerial survey camera, two video cameras, a laser profiling altimeter, and a PC based data recording system. Two aerial survey techniques for determining biomass are discussed. One primarily based on video and the other relying additionally on the 3D terrain models generated from the aerial photographs. In the first technique, transects are flown at 1,000 feet with dual-camera wide angle and zoom video, and a profiling laser operating at 238 Hz. The video coverage is used to identify individual tree species, and the laser profiler is used to estimate tree heights. The second procedure builds on this approach by taking sequences of 70mm photographs with an 80% overlap along a second higher altitude flight line at 4,000 feet. Detailed 3D terrain models are then generated from successive pairs of images. Several state-of-the-are computer vision algorithms are discussed, including the ITL system, which is an interactive ground cover classification system that allows an operator to quickly classify the large areas in a real-time, and Terrest, which is a highly robust 3D terrain modeling system. The work described in this paper is in a preliminary phase and all of the constituent technologies have not been fully integrated, we nevertheless demonstrated the value and feasibility of using computer vision techniques to solve environmental monitoring problems on a large scale.

System for real-time generation of georeferenced terrain models

A growing number of law enforcement applications, especially in the areas of border security, drug enforcement and anti- terrorism require high-resolution wide area surveillance from unmanned air vehicles. At the University of Massachusetts we are developing an aerial reconnaissance system capable of generating high resolution, geographically registered terrain models (in the form of a seamless mosaic) in real-time from a single down-looking digital video camera. The efficiency of the processing algorithms, as well as the simplicity of the hardware, will provide the user with the ability to produce and roam through stereoscopic geo-referenced mosaic images in real-time, and to automatically generate highly accurate 3D terrain models offline in a fraction of the time currently required by softcopy conventional photogrammetry systems. The system is organized around a set of integrated sensor and software components. The instrumentation package is comprised of several inexpensive commercial-off-the-shelf components, including a digital video camera, a differential GPS, and a 3-axis heading and reference system. At the heart of the system is a set of software tools for image registration, mosaic generation, geo-location and aircraft state vector recovery. Each process is designed to efficiently handle the data collected by the instrument package. Particular attention is given to minimizing geospatial errors at each stage, as well as modeling propagation of errors through the system. Preliminary results for an urban and forested scene are discussed in detail.

A self-consistency technique for fusing 3D information

This paper describes a robust method for recovering an optimal DEM and its variance from multiple, randomly orientated views of a surface. The method generates a set of DEM tiles in a common coordinate system from multiple overlapping images, and then employs the concept of self-consistency to detect and remove errors from the tiles. The clean tiles are averaged together to form a low noise composite DEM. The method is tested on real and photo realistic simulated data. Results show that the method is capable of producing a virtually error free composite DEM.

Scene reconstruction research — towards an automatic system

Current research towards three-dimensional reconstruction from aerial images at the University of Massachusetts is briefly summarized. The goal of this research is automatic robust reconstruction of both natural and cultural features under a variety of conditions and variable sensor data. We suggest that a promising direction for achieving this goal lies in the construction of systems containing context sensitive control strategies for synchronizing the application of image understanding modules (algorithms) whose individual domain of expertise is limited.