Construction material classification using multi-spectral terrestrial laser scanning

This research addresses the problem of populating Building Information Model databases with information on building construction materials using a new classification method which uses multi-spectral laser scanning intensity and geometry data. Research in multi-spectral laser scanning will open up a new era in survey and mapping; the 3D surface spectral response sensitive to the transmitted wavelengths could be derived day or night in complex environments using a single sensor. At the start of this research a commercial multi-spectral sensor did not exist, but a few prototype level instruments had been developed; this work wished to get ahead of the hardware development and assess capability and develop applications from multi-spectral laser scanning. These applications could include high density topographic surveying, seamless shallow water bathymetry, environmental modelling, urban surface mapping, or vegetative classification. This was achieved by using from multiple terrestrial laser scanners, each with a different laser wavelength. The fused data provided a spectral and geometric signature of each material which was subsequently classified using a supervised neural network. The multi-spectral data was created by precise co-positioning of the scanner optical centres and sub-centimetre registration using common sphere targets. A common point cloud, with reflected laser intensity values for each laser wavelength, was created from the data. The three intensity values for each point were then used as input to the classifier; ratios of the actual intensities were used to reduce the effect of range and incidence angle differences. Analysis of five classes of data showed that they were not linearly separable; an artificial neural network classifier was the chosen classifier has been shown to separate this type of data. The classifier training dataset was manually created from a small section of the original scan; five classes of building materials were selected for training. The performance of the classification was tested against a reference point cloud of the complete scene. The classifier was able to distinguish the chosen test classes with a mean rate of 84.9% and maximum for individual classes of 100%. The classes with the highest classification rate were brick, gravel and pavement. The success rate was found to be affected by several factors, among these the most significant, inter-scan registration, limitation on available wavelengths and the number of classes of material chosen. Additionally, a method which included a measure of texture through variations in intensity was tested successfully. This research presents a new method of classifying materials using multi-spectral laser scanning, a novel method for registering dissimilar point clouds from different scanners and an insight into the part played by laser speckle interpretation of reflected intensity