Automated Building Height Extraction and Building Detection from High Resolution Aerial and Space Imagery

analysis of the performance of this new algorithm for an urban and urban-rural boundary. An automatic height extraction algorithm using a stereoscopic approach and a building detection algorithm using a monoscopic approach are introduced. The auAUTOMATIC HEIGHT tomated height extraction is done using a pyramidal EXTRACTION matching strategy and a silhouette mask made from linHigh resolution urban area imagery have many disear elements of buildings. A pyramidal matching altinctive characteristics compared with low resolution imgorithm with a tile-based seed point selection strategy agery that make many approaches for low resolution between levels and automatic seed points a t the coarsimagery inappropriate. Buildings create abrupt height est level matching was developed. A silhouette image discontinuities, occlusions, and shadows. Moving vehiof buildings is used to remove possible blunders around cles are also detectable in urban area imagery. previous building boundaries. The accuracy of the algorithm is with the region growing G~~~~~~ adaptive assessed quantitatively and is shown to have 4m RMS. least squares correlation(ALSC) stereo matching dgoerrors for the area studied. rithm("Gotcha") [4, 51 shows that buildings make isoA novel building detection algorithm is proposed lated regions for region growing 161, the , , ~ ~ t ~ h ~ n based on a graph constructed from lines and line reneeds at least one initial seed point in an isolated region, lations. Building hypotheses are generated by finding a large number of seed points are required for urban area closed loops in the graph by a depth-first graph traversal stereo matching. algorithm. Building hypotheses are then verified after A pyramidal matching algorithm was implemented merging and removing false building hypotheses by usas a possible solution. Pyramidal matching solves the ing shadow lines and vertical lines. This algorithm was problem by utilising the results of lower resolution matchtested with 1 5 n resolution air-borne image and 2m resing , initial seed points in higher resolution matcholution space-borne image and shown to work successing. In this way, pyramidal matching facilitates well disfully. tributed initial seed points which cover as many isolated regions as ~ossible. Pyramidal matching also reduces INTRODUCTION the magnitude of height discontinuities and the number of isolated regions by averaging down the resolution of In the late 1980s an automated system for 3D ~ ~ ~ ~ d i ~ ~ t ~ the image. At the lowest (coarsest) level, matching can extraction, the UCL 3D I~~~~ Maker [I] was developed be performed successfully and in the higher level, matchfor 121 and spaceborne stereo 111. Several ing can be enhanced by utilising the results of lower level attempts were made at UCL [3] to enhance the 3DIM system for high resolution aerial photography includAnother advantage of pyramidal matching is the use ing the use of pyramidal matching and randomised seed of automatic seed points, which enables the developpoints. hi^ paper describes work on the use of "inment of an automated system. Automatic seed points teEgentn segmentation techniques to isolate regions of are generated assuming a disparity value blunders in stereo imagery as as a quanat the coarsest level and verified by non-region growing titative analysis of the accuracy of height measurement ALSC matching. As range at the coarsest level using automated stereo vision. will be small, actual conjugate points can be found af~ ~ t ~ ~ ~ t ~ d segmentation of buildings from aerial and ter ~re-matching. They are used as seed points to start high resolution (2m) Russian space imagery (DD5) uspyramidal matching. ing a line relation graph and a depth-first graph traverAn appropriate control strategy between levels should sal algorithm will be described as well as a quantitative be used to reduce the problem of blunder propagation. [6] In our pyramidal matching, a tile-based selection strategy was adopted as it can achieve well distributed seed points by defining tiles spread over the image space and selecting one match point with best accuracy within a tile as a seed point for the next level. The maximum eigenvalue of the covariance matrix of estimation is used as a measure of accuracy, which should be zero for an ideal matching. Strong edge elements presented in the imagery were used t o produce the silhouette image of buildings. The Canny-Petrou-Kittler(CPK) filter [7, 81 was used to extract edge elements. A Hough transformation based on Duda and Hart [9] and a connected edge labelling algorithm [6] were applied to the edge elements and their results were combined to select linear elements from building boundaries. A DEM of the linear elements was created by assigning a 3D coordinate to a point in a linear element based on the results from pyramidal matching. A silhouette image of buildings was then created by thresholding height to remove the ground level elements and applying a hole-filling operation to recover building structures from the wire-frame-like building boundaries. A Digital Elevation Model (DEM) from pyramidal matching was found to have high magnitude of errors near building boundaries. A silhouette image of buildings were used to overcome these errors. This is because it can locate buildings more precisely as it has been generated from linear elements. A silhouette image is used as a mask to a DEM from pyramidal matching to filter out the erroneous building level points. Fig l summarises the overall procedure.