Preliminary results on increased situational awareness for inland vessels using onboard LiDAR

This study discusses the preliminary results on increased situational awareness for inland vessels using onboard Light Detection and Ranging (LiDAR) technology on a 5 meter long research vessel. The pointcloud data set was captured on the Leuven-Dijle canal, over a 500m trajectory, resulting in an accumulated pointcloud of approximately 16 million points. This data set was first pre-processed, which involved cropping the pointcloud to a region of interest, voxel downsampling, and removing outliers. Next, as a first step towards localization and object detection, the shoreline corresponding to the trajectory is derived from this pre-processed data by means of 3 different methods: (i) a KDTree implementation followed by selecting the k-nearest points from a specific vessel location; (ii) a fitting approach using the shoreline derived from a raster image by means of image processing as a reference model; and (iii) a fitting approach using the shoreline derived from an Inland Electronic Navigation vector Chart (IENC) as a reference model. Subsequently, the resulting shorelines extracted from the LiDAR data set are compared with the IENC vector chart as this is currently the most accurate map available. The results indicate that the IENC vector fitting approach is the most desirable in terms of overall error. Nevertheless, this method, as well as the other fitting approaches, ignores points from potentially modified shorelines or vegetation growing within the shoreline area, i.e. the actual, dynamic environment is not captured. As such, this approach should merely be used for localization purposes or validating the IENC chart. The KDTree method, however, does account for these dynamic environments, and it seems to be capable of object detection at runtime. Furthermore, given that the results of the KDTree and IENC vector fitting approaches are very similar in terms of standard deviation and mean error (< 1m) between extracted shorelines and the IENC waterway features, the KDTree method is favourable for realtime use.