Automatic Detection of Buildings from Laser Scanner Data for Map Updating

Automatic interpretation of laser scanner data for building detection and map updating was studied. Digital surface and terrain models (DSM and DTM) and an intensity image were created using a helicopter-borne TopEye system recording with 2–3 pulses per m. The DSM was segmented into homogeneous regions using a region-based segmentation method, and the segments were classified as `building ́, `tree ́ or `ground surface ́. Height differences between the DSM and DTM, textural characteristics of the DSM and intensity image, and shapes of the segments were used in classification. Building segments were compared with an old building map and further classified as `new building ́, `enlarged building ́ or `old building ́. Similarly, building segments derived from the old map were classified as `detected ́, `partly detected ́ or `not detected ́. Comparison with an up-to-date building map shows that about 80% of all buildings in the study area were detected from the laser scanner data. For buildings larger than 200 m, the detection percentage was about 90%. Pixel by pixel comparison of the classification result with the reference map shows that 90% of pixels covered with buildings in the map were correctly classified. 85% of building pixels in the classification result were buildings in the reference map. The accuracy measures of the pixel-based comparison also include errors caused by small location differences between the data sources. According to visual evaluation, the most important changes between the laser scanner data and the old map, e.g. new buildings, were detected in change detection. The most problematic buildings for automatic detection were small buildings surrounded by trees.