Many technologies and applications now necessitate an awareness of their geographical surroundings, typically employing an array of sensors to capture the environment. An autonomous vehicle may utilise high definition cameras, LIDAR and RADAR; not only for collision avoidance but also for simultaneous location awareness and mapping (SLAM). Similarly, telecommunication network planning can benefit from the utilisation of RF propagation tools which include representations of target environments sourced from high resolution aerial photography and/or LIDAR point clouds. Mobile LIDAR scanners can capture a three dimensional environment and even colour each point using information from an associated camera. However, the resulting data is potentially very large, permutation invariant and clustered. Manually classifying this data, to maximise its utility in a propagation model, is not easily scalable; being both labour intensive and time consuming. This research defines a pipeline which facilitates the automatic classification of point cloud data and its subsequent consumption into a propagation model. It examines the use of sub-sampling as a mechanism to reduce the number of points to work with in order to reduce computational complexity and/or cater for lower resolution point clouds. It finds that a combination of edge-feature retention and sub-sampling at a resolution of 32 cm produces equivalent results to sub-sampling at 4 cm, whilst reducing the number of points by over 95%. Real-world LIDAR data, mapping 450 metres of an area in Bristol, UK, was captured, classified and converted into a series of simplified meshes before being input into a propagation model for comparison with human labelling and non-labelled (same class for all) data. Results superficially favour hand-labelled data, then automatically classified data followed by the non-labelled data. However, the results raise significant questions about whether the low-resolution of the hand-labelled data is truly representative of the real-world. In particular, how the coarseness of low-resolution can not only propagate rays more easily but also unintentionally block the line-of-sight. One major finding was the importance of labelling objects. This is not only to model the environment correctly, but also to discard undesirable objects which prevent propagation, if not removed.
Automated clutter classification of LIDAR data for Radio Frequency propagation modelling
Worsey, J. N. (Author). 2 Dec 2021
Student thesis: Doctoral Thesis › Doctor of Philosophy (PhD)