|dc.description.abstract||Terrestrial Laser Scanning (TLS) is a ground-based, active imaging method that rapidly acquires accurate, highly dense three-dimensional point cloud of object surfaces by laser range finding. For fully utilizing its benefits, developing a robust method to classify many objects of interests from huge amounts of laser point clouds is urgently required. However, classifying massive TLS data faces many challenges, such as complex urban scene, partial data acquisition from occlusion. To make an automatic, accurate and robust TLS data classification, we present a line-based multi-range asymmetric Conditional Random Field algorithm.
The first contribution is to propose a line-base TLS data classification method. In this thesis, we are interested in seven classes: building, roof, pedestrian road (PR), tree, low man-made object (LMO), vehicle road (VR), and low vegetation (LV). The line-based classification is implemented in each scan profile, which follows the line profiling nature of laser scanning mechanism.Ten conventional local classifiers are tested, including popular generative and discriminative classifiers, and experimental results validate that the line-based method can achieve satisfying classification performance. However, local classifiers implement labeling task on individual line independently of its neighborhood, the inference of which often suffers from similar local appearance across different object classes. The second contribution is to propose a multi-range asymmetric Conditional Random Field (maCRF) model, which uses object context as post-classification to improve the performance of a local generative classifier. The maCRF incorporates appearance, local smoothness constraint, and global scene layout regularity together into a probabilistic graphical model. The local smoothness enforces that lines in a local area to have the same class label, while scene layout favours an asymmetric regularity of spatial arrangement between different object classes within long-range, which is considered both in vertical (above-bellow relation) and horizontal (front-behind) directions. The asymmetric regularity allows capturing directional spatial arrangement between pairwise objects (e.g. it allows ground is lower than building, not vice-versa). The third contribution is to extend the maCRF model by adding across scan profile context, which is called Across scan profile Multi-range Asymmetric Conditional Random Field (amaCRF) model. Due to the sweeping nature of laser scanning, the sequentially acquired TLS data has strong spatial dependency, and the across scan profile context can provide more contextual information. The final contribution is to propose a sequential classification strategy. Along the sweeping direction of laser scanning, amaCRF models were sequentially constructed. By dynamically updating posterior probability of common scan profiles, contextual information propagates through adjacent scan profiles.||