BoreasLane: 3D Lane Dataset

BoreasLane is an initiative under the Toronto Robotics + AI Lab to create the first publicly available 3D lane dataset tailored for winter conditions. This project is part of WinTOR, a collaborative research program focused on advancing self-driving car technology in Canadian winters.

My contributions involved developing a high-performance pipeline for annotating the dataset by integrating sensor data from cameras, GPS, and LiDAR point clouds. Using techniques like caching, multiprocessing, and vectorized operations, I reduced preprocessing and labeling runtimes by 43%, enabling faster iterations and scalability.

The 3D lane labeling pipeline combines 2D lane annotations from satellite imagery with sensor data collected during vehicle drives. Lane markings are projected into the vehicle’s frame of reference and refined using LiDAR point clouds for high-accuracy 3D annotations. These annotations are exported in the OpenLane format to support model benchmarking.

Future work involves automating the entire data generation process, starting with unlabelled satellite imagery and leveraging HD maps for a 2D Bird’s Eye View Lane Detection Module. The ultimate goal is to publish the BoreasLane dataset alongside a research paper and to develop a Bayesian Attention Network-based 3D lane detection model, enabling uncertainty-aware segmentation for adverse weather conditions.