Drivable Area Segmentation on LiDAR Range View Images for Autonomous Driving

Autonomous driving is transforming modern mobility and redefining transportation paradigms. A fundamental requirement for such systems is the detection of drivable areas around the vehicle, which enables understanding of the surrounding environment and supports the motion-planning decisions that control the vehicle. This thesis tackles this problem by implementing a segmentation pipeline to identify road regions that the vehicle can traverse, with the goal of providing a detailed representation of the road structure and its components. The central focus of this work is the segmentation of drivable areas and road markings using a neural network–based approach. The main novelty lies in the use of Range View images generated by a LiDAR sensor, which significantly differ from conventional camera images. This design choice is motivated by the stringent real-time constraints of autonomous driving, necessary to build highly responsive systems. The resulting segmentation model delivers encouraging performance in terms of both segmentation quality and inference speed. Because no LiDAR-based datasets for road segmentation are available, both training and evaluation are based both on the BDD100K dataset and on a proprietary dataset developed specifically for this thesis. To construct it, a hybrid annotation pipeline is introduced, combining automated and manual labeling methods to efficiently produce a high-quality annotated dataset. The annotation statistics gathered during this process highlight the effectiveness of this semi-automatic strategy. The final element of the proposed segmentation pipeline is a grid map that applies a perspective transformation to the model's outputs, converting the front-facing view into a bird’s-eye-view representation. Road-related classes detected by the model are then aggregated over consecutive frames, resulting in a more complete depiction of the vehicle’s surroundings that remains robust in the presence of occasional segmentation outliers.

Author dc:creator

• Simone Bevilacqua (24399026)

• Autonomous Driving
• Deep Learning
• LiDAR
• Semantic Segmentation
• Drivable Area
• Road Line
• Grid Map

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.32991911.v1

OAI identifier oai:identifier

oai:figshare.com:article/32991911