Racing dataset boosts 3D lane detection to 300Hz with odometry fusion

Lane detection for autonomous driving has primarily focused on regular road conditions, but high-speed racing introduces extreme geometries and rapid motion. To address this, researchers from Toyota Research Institute and Lund University created a new dataset called the Racing Lane Dataset (RLD), capturing over 250,000 images from multiple cameras and inertial measurement units (IMUs) mounted on a Lexus LC 500 driving on a closed circuit. The dataset includes synchronized odometry data, enabling a novel approach that fuses visual lane markings with pre-integrated inertial measurements over time.

Their method processes frames at nearly 300Hz — far faster than typical autonomous driving systems — while retaining high accuracy. By incorporating odometry and ensemble predictions from multiple cameras, the system achieved an F1 score above 0.9 and lateral mean absolute error (MAE) under 0.18 meters during real vehicle deployments. Compared to the state-of-the-art BevLaneDet, this approach improved F1 by 3 points and cut near-vehicle MAE by more than 30%. The work, accepted as an IROS paper, demonstrates that inertial measurements can effectively regress road geometry across both spatial and temporal dimensions, making it practical for racing scenarios where every millisecond matters.