Evaluation of several software packages in context of point cloud classification
The national survey agency is responsible for generating nation-wide digital terrain and surface models (DTM and DSM). For this reason, they are gathering Airborne Laser Scanning (ALS) data of their area every three years, which have a rather coarse resolution of 4~10 points/m². At the same time, they derive point clouds from optical image data using a method called ‘Dense Image Matching’ (DIM) resulting in point clouds with a resolution of 100 points/m². The goal of this project is to evaluate established software tools for both data types (ALS and DIM). The data sets have to be classified using ArcGIS software as well as an approach proposed by Maltezos and Ioannidis (2015). In this paper, they use a combination of cloud compare functions to calculate normal and roughness features of the point clouds as well as some written code, which utilizes those features for classification. Lastly, both classification methods on both data sets have to be evaluated using manually classified reference point clouds.
Processing of large-scale data sets in the context of autonomous driving
The research training group i.c.sens has produced large quantities of data to support scientific research in the context of autonomous driving. To this end, multiple cars have been equipped with complex sensor setups for self-localization and mapping, including multiple GNSS systems, stereo cameras and multiple LiDAR systems. In order to enable secondary usage of these data sets and to publish the data set at a later point in time, the data needs to be prepared using established sensor-specific data processing methods or manual data annotation processes (e.g. labeling of images or point clouds towards a reliable ground-truth). The range of possible activities (programming, using a GIS for analyses, manual editing/annotation of data using provided tools and many more) in this research project is wide and can support multiple students at the same time. However, there is a lot of flexibility in determining the specific tasks to carry out (in a meeting before starting the research project).
Extracting Relevant Features That Determine Collision Avoidance in Shared Spaces
DFG Graduiertenkolleg SocialCars
In distinction to classic traffic designs which, in general, separately dedicate road resources to road users by time or space division, an alternative solution—shared space—has been proposed by traffic engineers. Pedestrians, cyclists, and vehicles interact with each other and self-organize to give or take right-of-way. The safety of shared spaces need to be thoroughly investigated, namely, how road users adapt their speed and/or orientation in the interactions with others in their vicinity to avoid collisions. In order to extract the most relevant features that reflect how a road user adjust his/her motion to avoid potential collisions with others in shared spaces, real-world trajectories will be analysed using statistical and machine learning approaches. For instance, the safe distance may differ significantly across different types of road users. Can we quantify such differences and impacts? Currently, however, user attributes are not yet available in the dataset, which will be incorporated in future work.