Optimizing labelling on radar-based grid maps using active learning

The aim of this work is to optimize labelling on radar-based grid maps. Therefore, we suggest an active learning system where only the most valuable samples in the training set are to be labelled manually. We show that this approach drastically reduces the overall time needed for labelling. Using only about 40% of the samples in the training set yet still leading to the same classification results as the completely supervised reference experiment, the proposed approach is suited for optimized dataset creation in the given data domain.