Detection and Recognition of Objects from Mobile Laser Scanning Point Clouds: Case studies in a road environment and power line corridor

Accurate and detailed three-dimensional (3D) geospatial data are needed, and they open new future possibilities. The collection and processing of data should be efficient to add value for society. Mobile laser scanning (MLS) is a detailed 3D mapping technique that can produce 3D point cloud data with a sub-centimetre relative accuracy. Data collection is efficient, and in a road environment, hundreds of kilometres can be mapped daily. MLS produces large amounts of data (eg 1 gigabyte per 1 km of road), and the processing should be automated for the technique to be practical and efficient. This thesis studies the automated interpretation of the 3D point clouds collected using MLS. First, it investigates how completely and correctly polelike objects, such as traffic signs and lampposts, can be detected from MLS point clouds in a road environment. Second, it studies if the accuracy of general object recognition from MLS point clouds can be improved using new algorithms. Third, the detection of power line components is studied outside the road network, both in forests and on farmland. Object detection and recognition algorithms are developed and tested with real-world data collected in a road environment and a power line corridor outside the road network. The algorithms' accuracies are evaluated quantitatively. The results suggest that polelike objects can be detected with an accuracy of between 70% and 87% in a suburban environment. These results are among the first to evaluate the accuracy of polelike object detection quantitatively, implying slightly higher accuracies than previous studies. For the first time, local descriptor histograms (LDHs) are applied to machine-learning-based object recognition from MLS point clouds in surveying applications. The results suggest that LDHs can increase the accuracy of the point cloud segment classification – a 9.6 percentage point increase is observed compared to a state-of-the-art accuracy of 78.3%. Undersegmentation and incomplete ground extraction are the most prominent error sources in object recognition. MLS is applied for the first time in power line detection outside the road network, in forests and on farmland. The results imply that the accuracy of automated power line detection may be higher than 93%. Most errors are caused by the inaccuracy of the trajectory and attitude determination, probably due to tall trees blocking navigation satellite signals. This dissertation demonstrates the potential of MLS in the automated mapping of the road environment and the power line corridor outside the road network. The thesis also presents techniques that may improve the accuracy of the automated data interpretation. In addition, the dissertation points out some limitations of MLS technology through error analyses, providing directions for future research.