Building Detection from LiDAR PDF Print E-mail
Written by Nicholas Shorter   
Monday, 04 August 2008 05:05

Why develop a building detection algorithm in the first place? The primary motivations for developing the proposed building detection algorithm are as follows: to derive an automated technique capable of individually isolating buildings from irregularly spaced Light Detection and Ranging (LiDAR) data; then having the algorithm extract those isolated, individual buildings; and finally presenting them to a 3D reconstruction algorithm.

Building Detection Block DiagramThe buildings are detected, isolated and extracted from the raw, irregularly spaced LiDAR data (not interpolated/rasterized LiDAR data). It is emphasized that the algorithm operates on the raw LiDAR data as interpolating/rasterizing the data to fixed points (subsequently creating a range image) introduces unwanted noise / inaccuracies and other side effects.

What other methodologies exist for building detection? What advantages does the proposed building detection method have over existing methods? Several methodologies existent in the literature make use of morphological filters to differentiate between terrain and non-terrain points. Zhang et. al. note that the problem with a static window size, for morphological operations, is the ideal window size is not always the same size throughout the duration of a given single data set. Therefore the window size in must be tuned to the data set, if an optimal window even exists. One must ask then, why even bother using a window for building detection? Would it not be optimal then to have an algorithm that did not rely on a window size? The proposed algorithm does not use a windowing technique; rather it instead analyzes features from the irregular spaced LiDAR triangulation to make its decision on how to differentiate building from non building points.

 

 

 

 

Two scenes covered by aerial images are presented as follows:

Scene 1

Scene 2

Each building in the below scene is represented as a unique color while the ground is represented as a dark blue. Some of the colors may appear very similar in hue to the dark blue ground color; however in this scene the majority of the buildings were successfully detected (96% of them to be exact). Matlab’s (the temporary solution chosen to render the output of the algorithm) automatic coloring scheme will assign groups relatively close in number label to a similar hue. Note: Matlab only plots the end result, the triangulation and building detection were custom developed algorithms (not built in functions).

 

Building Detection Output

 

In the below output, buildings are labeled as dark red whereas all non-building entities are labeled as light green:

 

Building Dection Output 2

 

On Friday, October 30, Nicholas was invited by Dr. Georgios Anagnostopulos to give a lecture at the Florida Institute of Technology (FIT) about his dissertation research, specifically his work with unsupervised building detection from Irregular LiDAR data.

More information about this lecture has been provided here.

 

Fore more on this research effort please refer to the following conference paper and poster presentation:

 

Shorter, N. S.; Kasparis, T. (2008). Triangulated, Connected Sets for Building Detection from Irregularly Spaced LiDAR. The 3rd International Symposium on Communications, Control and Signal Processing, St. Julians, Malta, March 12 – 14 (published)

Conference Paper

Posterboard for Triangulated, Connected Sets for Building Detection from Irregularly Spaced LiDAR presented at The 3rd International Symposium on Communications, Control and Signal Processing

Conference Poster

Last Updated on Friday, 18 December 2009 19:47