Segmentation of laser-altimeter data for building reconstruction: Comparison of different procedures.


Motivation. Altimeter data recorded by airborne laser scanners are gaining increasing importance in the generation of three dimensional city models. There is a specific request for such models in the visualisation of urban scenes, in mission planning, in disaster management and as a base for simulations, e.g. the areas of environmental sciences and telecommunication. A main component of a city model is the vector description of the buildings with their roofs. For our analysis of the roof shape the elevation data of each building is masked using a digital map. In the roof reconstruction the segmentation of the elevation data is the first processing step and plays an important role. For the segmentation of range data several procedures have already been developed. In previous publications procedures were investigated for handling laser scan images and structured light images of some polyhedral objects in a laboratory like environment. An open question is which of these procedures is most adequate for the task presented here and the special data of natural scenes with artefacts. Methods. In this contribution four range image segmentation procedures known from various publications are investigated and compared. Three procedures were applied which are based on region growing. Another procedure for the extraction of straight lines also produces a segmentation into planes as an intermediate result. Furthermore a simple procedure was developed, which is controlled by only one parameter and which, in addition, yields the neighbourhood graph of the segments. Its distance measure is not produced from derived properties of segments but directly from their point sets. Each of the procedures produces a label image and, for each segment, a description of the plane by the parameters of its equation. For each of the procedures experiments with extensive parameter variation were carried out. Ground truth was determined manually from the elevation data. For the valuation of a single segment the intersection between the segment and its ground truth is used. For the comparison of whole segmentations several aspects of the single segment valuation are discussed and a whole segmentation valuation function is developed. Results of applying the segmentation valuation function are demonstrated using selected buildings. Results. In carrying out the experiments, the determination of optimal parameters proved difficult. In some procedures this resulted from the large number of parameters and interdependencies of different parameters. In general the low horizontal scale of 1m led to difficulties in segmenting details of buildings such as dormers or small extensions. The same image data were rotated by 90, 180 and 270 degrees. Under these rotations each of the investigated procedures showed some changes in the segmentation result. The procedure which determines a partition based only on the gradient orientation showed the least variation under these rotations. However this procedure was sensitive with respect to noise contained in the images. Some of the procedures investigated assume different sensor specific geometry which is not present in the data used. In one of the procedures the assumed line-like sensing structure led to a stripe-like oversegmentation. The procedure which produced the best results also needed most computation time. Conclusion. The valuation function was built with subjectively chosen weights. For example, oversegmentation was tolerated more than undersegmentation. Depending on the problem posed and the intended further operation, other segmentation procedures may be optimal if different valuation functions are used. Since the investigation was done using the data of only one flight, data sets of further flights at different seasons and with different sensors need to be inspected for a validation of the obtained results. In order to estimate the problems caused by segmentation faults in the reconstruction of total buildings, the vector descriptions produced later should also be evaluated by comparing to a ground truth in a test bed.
Geibel R, Stilla U (2000) Segmentation of laser-altimeter data for building reconstruction: Comparison of different procedures. International Archives of Photogrammetry and Remote Sensing. Vol. 33, Part B3, 326-334
[ Stilla.de/pub ]