Segmentation of floating objects in LIDAR dataMirčeta, Kristijan (Avtor) Marolt, Matija (Mentor) Bohak, Ciril (Komentor) computer visionmachine perceptionsegmentationpoint cloud segmentationairborne LiDAR simulationLiDAR augmentationIn this thesis we attempt to segment floating objects – for example, airplanes, birds, balloons, and drones – in airborne LiDAR data. We present a heuristic method (RBNN) for discovering floating objects and evaluate it on a subset of the ARSO dataset, which is a scan of the entire terrain of Slovenia. The results offer many candidates but only one true floating object in 3 square kilometers of terrain. We conclude that there are too few positives to solve this problem via supervised learning. To address this issue, we augment the dataset with virtual floating objects and then train a machine learning model to find these in the augmented dataset. The idea is that such a model will then be able to find similar objects in unseen data that are in fact real floating objects. We present an augmentation pipeline with which we can augment any LiDAR dataset with any polygonal model (known as an augmentable). We determine the quality of the augmentation by evaluating the consistency of the virtual scanning with the real scanning. This is composed of two subproblems: where to place the augmentable and how to scan it. Our method reliably solves the problem of placing the augmentable. The difficulty in scanning lies in determining the scan direction and the location of the scanner during scanning. We determine the direction with 18 angular degrees of error, computed on 2,219 augmentables, over 50 square kilometers of ARSO dataset terrain, and we can reliably determine the location of the scanner if we know its direction. In the end, we use a modern point cloud segmentation method with the augmented data. The algorithm has advantages and disadvantages when compared to RBNN, and it may be beneficial to combine both methods.20192019-12-16 15:00:02Magistrsko delo/naloga113243sl