Classıfıcatıon of Poınt Clouds Acquıred Through Mobıle Laser Scanner in Urban Areas Usıng Geometrıc and Shape Features
Özet
The mobile laser scanners (MLS) serve as a high density, high accurate and faster data collection method for urban areas from a street-level surveying perspective. The major processing phases of a MLS point cloud classification can be considered as (i) the neighborhood selection, (ii) the feature extraction, and finally (iii) the classification. Since MLS point clouds are poor in terms of the attributes, the classification phase must be supported by using features derived from the local neighborhood relations between points in a dataset.
This thesis deals with the point-based supervised classification of point clouds acquired through a vehicle-based MLS system in urban areas using local geometric and shape features. The developed approaches are tested using the benchmark dataset representing the Technical University of Munich (TUM) City Campus. The local features of each point in the point cloud were extracted and evaluated through three different neighborhood definitions, i.e. spherical, cylindrical and the k-nearest neighbor. As the classification strategy, the Random Forest (RF) classifier that has been preferred in quite a few studies dealing with MLS classification is applied, and is successively tested for the point-based supervised classification. A total of 8 classes are involved during the classification: artificial terrain, natural terrain, high vegetation, low vegetation, building, hardscape, artifact and vehicle. The results were evaluated as classification for all three local neighborhood types with different parameters, and for various combinations of features. The results achieved were compared with three previous studies utilizing the same data set in the literature, and a combination of the features from different neighborhood information increased the overall results at least 4% with considerable improvements (up to 40%) for the producer’s accuracies of multiple classes.
