3D Lane Detection with Odometry for High-Speed Vehicle Racing

Autonomous vehicle research has traditionally focused on urban driving, where speeds are moderate and road geometries are predictable. Vehicle racing presents a significantly more challenging environment characterized by extreme road geometries, high speeds, and dynamic maneuvers. Existing 3D lane detection models are often optimized for static, monocular urban driving and struggle to maintain the high inference rates and geometric accuracy required for racing applications.

To address this gap, the authors introduce RaceLane, a comprehensive dataset featuring over 250,000 images captured from a Lexus LC 500 on a closed circuit. The dataset includes multi-camera feeds, inertial measurement unit (IMU) data, and wheel odometry.

The authors propose ENSEMBLELANES, a modular framework that enhances existing 3D lane detectors. Key innovations include:

• Road-Frame Alignment: Aligning the Bird's Eye View (BEV) space with the vehicle's road frame to simplify regression and improve recall.
• PCA-based Clustering: Replacing computationally expensive clustering algorithms with a vectorized Principal Component Analysis (PCA) approach, enabling inference rates exceeding 290Hz.
• Temporal and Multi-Camera Fusion: Utilizing IMU and odometry data to perform on-manifold pre-integration, allowing the system to regress road geometry by fusing asynchronous predictions across both time and multiple camera angles.

By integrating odometry and ensemble predictions, the authors demonstrate a 3-point improvement in F1 score and a reduction of over 30% in near-vehicle mean absolute errors (MAEs) compared to baseline methods like BevLaneDet. The system achieves F1 scores greater than 0.9 and lateral MAEs of less than 0.18m in real-world vehicle deployments, proving that leveraging ubiquitous vehicle sensors can significantly boost the robustness of 3D lane detection in high-speed, dynamic environments.