What? Why? How? PDF Print E-mail
Written by Nicholas Shorter   
Tuesday, 25 March 2008 06:04

What is the 3D Reconstruction problem as it applies to Geological Information Systems?

One could model 3D Reconstruction as a black box which operates on a given collection of inputs, inputs which depict a terrain (in this case with building structures on that terrain). The black box, extracting and analyzing key features from those inputs, then produces a 3D representation of the terrain. The inputs considered for that black box for this research are a point cloud captured by a LiDAR sensor and an aerial image. The output of the algorithm, for this research, will be 3-Dimensional models with the aerial image mapped onto those models. It is important to note that the 3-Dimensional models are not simply a triangulated mesh or an interpolated surface, but rather models completely described by geometric features, which represent the buildings existent in the terrain. For example a rectangular building is represented by 3 rectangular planes which are connected to one another having common sides, and the ends of the walls of that terrain resting on the bare earth surface. For this research, an emphasis will be placed on completely automating the execution of the 3D reconstruction algorithm (no user intervention).

Why would anyone need 3-Dimensional models of building structures existent in a given terrain?

Applications of 3D Reconstruction have valued use for both militaristic and commercial purposes. For the military, the analysis of LiDAR data can be used for target recognition applications. Work in training volume correlation filters (VCFs) to recognize tanks and other military vehicles within LiDAR data has recently been developed [3]. There is investigation underway for mounting LiDAR sensors on unmanned aerial vehicles (UAVs) [5]. This would enable aerial surveying of terrains in which military forces were denied access too. Scenes surveyed by an UAV or high flying plane with a LiDAR sensor could then be reconstructed into 3D models.

For commercial uses, the demand for 3D models of buildings has applications such as urban planning, network planning for mobile communication, spatial analysis of air pollution and noise nuisances, microclimate investigations, geographical information systems, and security services. For entertainment purposes, 3D reconstruction can be used for tourism information systems. Tourists, instead of using 2D maps to find their way around theme parks, could use a kiosk to view a virtual 3D walk through of the park. Also, 3D Reconstruction has change detection applications [4], where buildings demolished by natural disasters, are automatically identified via histogram difference thresholding comparing a pre and post event LiDAR analysis.

What is a point cloud and how is it captured?

Mounted on the aircraft, helicopter or plane, collecting the LIDAR data, is a Global Positioning System (GPS), an Inertial Navigation System (INS) and a LIDAR sensor system. The GPS returns the longitude and latitude coordinates of the aircraft’s current position. The INS tracks the altitude of the LIDAR sensor. The LIDAR sensor system is an active remote sensing instrument which consists of an emitter and receiver. The emitter sends out a pulse of light (electromagnetic radiation) into the atmosphere. The telescope (receiver) measures the intensity of the signal scattered back to the sensor system after that signal has interacted with various constituents in the atmosphere [1],[2]. The time from departure to return is also recorded thus enabling the calculation of the distance of the sensor to the target (or in this case the sampled terrain). The information received independently from the collection of these sensors culminates in the longitude, latitude and elevation (or (X,Y,Z) Cartesian coordinates) for a given point in the terrain sampled by the sensors. That information is collected for thousands of points for a given terrain and makes up a point cloud file. Therefore, a point cloud file is typically an ASCII file containing longitude, latitude and elevation values for several thousand points in a given terrain.

What benefits come from automating the 3D Reconstruction algorithm?

A given ASCII point cloud data set can contain on the order of millions of points and trillions of characters. One can purchase LiDAR from a commercial company for approximately a thousand dollars per square kilometer. Therefore a single square kilometer could contain hundreds of buildings. Consider a company having hundreds of data sets consequently containing tens of thousands of buildings. If a reconstruction algorithm has parameters associated with its execution for each data set, optimizing these parameters for a given data set could prove tedious. Furthermore, if user intervention is necessary during the reconstruction process for complicated buildings, this would result in having, typically, a trained technician on site to help with the algorithm’s execution for what could be hundreds to thousands of complicated buildings - hardly an efficient process. One facet of my research is to emphasize complete automation so that the reconstruction algorithm works as follows: the algorithm reads in the data set; the algorithm automatically reconstructs 3D building models with aerial images automatically mapped onto those models; the models are then returned to the user.

Why is not possible to reconstruct perfect representations of the sampled terrain?

There are several limitations which restrict perfect reconstruction of the terrain sampled by LiDAR sensors. These limitations result from several phenomena: noise introduced from multiple stages of the LiDAR sensing process and limited sensor measurement accuracy. Filin reports that because the geolocation of a single point results from the integration of three subsystems (GPS, INS, and LiDAR), it is possible that errors can come from any combination of all three sensors [6]. Specifically, errors resulting from the three subsystems include but are not limited to the following: a constant offset in the range determination, inaccurate scan angle determination, mounting bias from the misalignment between the INS and LiDAR sensor, GPS offset and drift, and INS system drift. These reported errors only exist as systematic discrepancies between the integration and limitations of the sensors themselves. As the electromagnetic radiation that is the laser is propagated through the atmosphere both its intensity and path are distorted by the interaction between the beam and the atmosphere itself [1],[2]. An additional aspect in which influences the LiDAR procurement process is artifacts resulting from scan angles not uniformly interacting with the sampled terrain due to obstructions from the terrain itself. Consider as an aircraft flies adjacent to a tall skyscraper, the laser pulses will interact with the side of that tall skyscraper but not with the terrain behind it resulting in a shadowing effect [4].

What sources can and/or should be used to generate 3D models of structures existent in a terrain? Why limit the algorithm to only using a point cloud and single aerial image file for inputs to the algorithm? Why not consider additional inputs?

A multitude of potential sources exist for 3D reconstruction: LiDAR point cloud, a single overhead aerial image, multiple aerial images from different perspectives, an overhead video sequence, 3D GIS ground plans, existing 3D GIS models, and infrared images. One potential issue with 3D GIS ground plans and existing 3D GIS models is overtime, as new buildings are created and old ones demolished, the plans and models become outdated. Furthermore, these models and plans only exist for select areas. Also, multiple images from varying perspectives, infrared images, as well as video sequences are uncommon. However, an overhead aerial image and LiDAR point cloud are either readily accessible sources of data or can be procured. Therefore the developed algorithm for my research will only consider the following two input sources: LiDAR Point Cloud of at least 1.5 points per m^2 point spacing density and an overlapping aerial image depicting the same scene as the point cloud.

What work have you done related to 3D reconstruction? How do you aim to accomplish the above mentioned goals?

All related publications can be found in the ‘Works Completed’ portion of my website. Some of the strategies I have planned for my research to accomplish the above mentioned goals have been outlined in my ‘PhD Candidacy Presentation’. If you have any additional questions feel free to contact me at This e-mail address is being protected from spambots. You need JavaScript enabled to view it

 

References:

[1]

Sun, Bing Yu; Huang, De-Shang; Fang, Hai-Tao; “LiDAR Signal Denoising Using Least Square Support Vector Machine.” IEEE Signal Processing Letters, Vol. 12, No. 2, February 2005

[2]

Yu, Shirong; Wang, Weiran; “LiDAR Signal Denoising Based on Wavelet Domain Spatial Filtering.” International Conference on Radar, October 2006

[3]

Mahalanobis, Abhijit; “Multidimensional Algorithms for Target Detection in LiDAR Imagery.” University of Central Florida, Electrical and Computer Engineering Seminar Series. Orlando. 28 March 2007

[4]

Vu, Tuong Thuy; Matsoka, Matashi; Yamazaki, Fumio; “LiDAR based Change Detection of Buildings in Dense Urban Areas.” Geoscience and Remote Sensing Symposium, 2004. IGARSS '04. Proceedings. 2004 IEEE International

[5]

Lammons, George; “After The Storm.” Military Geospatial Technology, Volume 4, Issue 1, March 2006.

[6]

Filin, Sagi; “Elimination of Systematic Errors From Airborne Laser Scanning Data.” Geoscience and Remote Sensing Symposium, July 2005.

 

Last Updated on Friday, 03 October 2008 17:26