PhD Defence Mr Adam Patrick Nyaruhuma
Department of Earth Observation Science
Title of defence
Automatic verification of buildings using oblique airborne images
The world has recently witnessed fast increase in systematic acquisition of airborne oblique images with scenes taken from multiple directions. Companies with large datasets include Pictometry (and their licensees Blom) and Slagboom en Peeters Luchtfotografie B.V. More interesting are continuous announcements of development of new camera systems including the Microsoft’s Osprey, the Leica RCD30 5-head, the IGI Penta DigCam and the rotating MapVision’s A3 camera.
This research investigated the usefulness of oblique airborne images for automatic verification of buildings in topographic datasets. Existing building verification methods utilize colour, texture and height from vertical images or range data. Oblique images, which contain top as well as side views of imaged objects, are not commonly used.
In this work, two methods were developed, the first for verification of building outlines in two-dimensional (2D) large scale topographic databases and the second for verification of 3D building models. In both methods, the vector data was fitted to multiple oblique images and features derived from the images were used to do the verification.
For 2D building verification, the 2D outlines were converted to 3D wall hypotheses and then features based on edges and texture of walls in multi-perspective images were automatically computed and compared. In this case, a number of issues were noted as important. Firstly, the results have shown that the clues identified and modelled to obtain verification measures were useful for building verification.
A second issue noted as important is the approach for combining the different measures for possible existence of a wall in images. In the developed method, verification of 2D building outlines starts by checking individual walls of a building and then the results are combined for overall verification of the building. A number of features suitable for recognising a wall in oblique images were therefore developed and strategies to combine these features in an overall measure of the status of a wall and a building were designed and tested. In our tests for combining wall verification measures with Adaptive Boosting, Random Trees and a variant of the Dempster-Shafer (Hints), the results showed minor differences but in most cases Adaptive Boost produced the best results.
The third aspect relates to the need for identification of occluded (sides of) buildings which would otherwise be confused with demolished ones. For this purpose, our procedure for visibility analysis proved to be very useful. The procedure uses a point cloud generated from the same oblique images, thanks to already existing image matching methods.
Regarding the evaluation, the completeness and correctness obtained (around 90 - 100%) was very encouraging. As the method verifies individual walls and then combines the results to obtain the per-building status, the method identifies not only completely existing or demolished buildings but also buildings that may have changed by part demolition or extension. These are signalled when only some walls of a building are identified in the images.
For the 3D case, the problem being tackled is identifying buildings that are demolished or changed since the models were constructed or identifying wrongly constructed models using the images. Automatic method for verification was developed by adopting the theory of Mutual Information. It assumes that the pixel gradient directions computed along a model edge should be generally different from gradient directions computed on random image positions. These gradient directions were found to be very robust for the verification.
Some important issues were also noted. Firstly, by concentrating on roofs only we also used the method to test and compare results from nadir images. This comparison made clear that especially height errors in models can be more reliably detected in oblique images than in nadir images because of the tilted view.
The second aspect relates to spatial resolution and overlap of image. Tests with images of higher overlap and resolution (Slagboom en Peeters) gave more mutual information but the mutual information in the relatively lower resolution (Pictometry) images was already sufficient and the verification results from the two types of images were similar. The correctness and completeness were between 97% and 100% in both cases.
The third issue is related to the parts of a building that are used for verification. The combining roof and wall information proved to be more useful than when using only roof or wall information alone.
Although complete and general automation of building extraction is still challenging we have shown that good results can be achieved for the subtask of automatic verification of existing datasets.
Adam Patrick Nyaruhuma was born in Muleba, Tanzania on 5th February, 1971. He did a bachelors’ degree at Ardhi University and worked for the Ministry of Lands, Housing and Human Settlements Development in Dar Es Salaam, Tanzania.
In 2007 he attained an MSc in Geo-informatics from ITC, Enschede, the Netherlands. His MSc. research was on performance analysis of algorithms for detecting roof faces in airborne laser scanner data.
Upon successful completion of the MSc, he got an opportunity, in 2008, to do PhD study at the University of Twente, The Netherlands. The research was on using oblique airborne images for automatic verification of buildings in topographic datasets.
Nyaruhuma, A.P., Vosselman, G. (Promotor) , Gerke, M. (assistant promotor) and Mtalo, E.G. (assistant promotor) (2013) Automatic verification of buildings using oblique airborne images. Enschede, University of Twente Faculty of Geo-Information and Earth Observation (ITC), 2013. ITC Dissertation 6235, ISBN: 978-90-6164-364-7.
|Event starts:||Thursday 07 November 2013 at 16:30|
|Venue:||UT Waaier 4|
|City where event takes place:||Enschede|
|Country where event takes place:||Netherlands|