Adverse Weather Benchmark Dataset for LiDAR-based 3D Object Recognition and Segmentation in Autonomous Driving

Current developments in flexible mobility solutions are striving towards autonomous electric driving in the areas of public transport and logistics. The benefits of such systems are lower costs, higher availability, and greater flexibility. Object detection and segmentation techniques based on LiDAR sensors to complement camera and GPS data are essential for reliable behavior in autonomous driving. However, little research has been done to evaluate these techniques in representative adverse weather conditions such as rain, fog, or snow. Consequently, this paper presents a new dataset based on adverse weather data present in already widely used public datasets. The existing data is complemented with additional weather labels to facilitate the evaluation of object detection and segmentation in various weather conditions. To generate a baseline, a state-of-the art 3D object recognition is evaluated using the enhanced dataset. The results show a strong impact of the weather conditions on the performance of the evaluated baseline algorithm, indicating the relevance of the benchmark.