Abstract
Accurate localization for autonomous vehicles relies heavily on the effective association of sensor-detected lane markings with pre-existing map landmarks. Traditional representations using 2D points or polylines often lack sufficient geometric context, leading to failures in high-noise environments or ambiguous road segments. This paper introduces the Delta-Angle Lane Marking Representation (DA-LMR), which augments spatial coordinates with local curvature information. We further propose Distance-Compatible Sample Consensus (DC-SAC), a heuristic refinement of RANSAC that leverages geometric constraints to accelerate the association process. Our evaluation on synthetic and real-world datasets demonstrates that DA-LMR provides superior robustness to sensor noise, maintaining high precision where traditional methods degrade.
Introduction
For an autonomous vehicle to navigate safely, it must resolve its position relative to a map with high fidelity. This process, known as data association, involves matching real-time sensor detections to known landmarks. While significant research focuses on the optimization back-ends of localization (e.g., SLAM), the underlying representation of the data—the "what" of the association—is frequently overlooked.
Standard systems typically represent lanes as 2D points $(x, y)$ or polylines. However, stripping a road's geometry down to simple coordinates discards valuable context. In noisy environments or areas with repetitive features, these simple representations struggle to distinguish between distinct road segments. We address this by encoding the "shape" of the lane directly into the data points, creating a more unique and robust signature for matching.
Methodology
Delta-Angle Lane Marking Representation (DA-LMR)
The core innovation of our approach is the transition from a 2D representation to a 3D feature space. Instead of a simple coordinate, each lane point is represented as $(x, y, \Delta\alpha \cdot w)$. Here, $\Delta\alpha$ (the "delta-angle") represents the change in orientation between consecutive line segments, effectively capturing the local curvature of the road.
A critical property of DA-LMR is its direction invariance. Whether a vehicle traverses a lane in one direction or the opposite, the calculated curvature remains identical. This consistency is vital for robust mapping and localization across different driving passes.
Distance-Compatible Sample Consensus (DC-SAC)
To perform the actual matching, we introduce DC-SAC, an optimized version of the RANSAC algorithm. Standard RANSAC is computationally intensive as it evaluates random pairs of points. DC-SAC prunes this search space by enforcing geometric consistency: it only considers pairs of detections whose Euclidean distance matches the distance between potential map landmarks within a predefined threshold $\gamma$. This heuristic significantly reduces the number of hypotheses to test, improving both speed and accuracy.
Experimental Evaluation
We validated our approach using a combination of synthetic and real-world data:
- Synthetic Dataset: Based on a hand-labeled map of Karlsruhe, Germany. We simulated sensor inaccuracies by adding Gaussian noise (0.1m to 0.5m standard deviation) and 10% outliers.
- Real-World Data: Qualitative testing was performed using the BerthaOne experimental vehicle, equipped with stereo cameras, to ensure performance in diverse driving conditions.
Results and Discussion
DA-LMR and DC-SAC significantly outperformed traditional Point Representation (PR) and Line Segment Representation (LSR). The primary advantage of DA-LMR is its stability; while LSR can achieve high precision in low-noise settings, its performance collapses as sensor jitter increases.
In high-noise scenarios ($\sigma = 0.5m$), our method achieved 98.1% precision and 99.7% recall. By treating curvature as a unique identifier, the algorithm effectively filters out false positives that plague 2D-only methods. Furthermore, DC-SAC demonstrated computational efficiency suitable for real-time onboard processing, matching RANSAC's speed while providing superior localization results.
Conclusion
By moving beyond simple $(x, y)$ coordinates and incorporating local curvature, DA-LMR provides a more resilient foundation for autonomous vehicle localization. The combination of this unique representation with the geometrically-constrained DC-SAC algorithm offers a robust solution to the data association problem in challenging, noisy environments. Future work will focus on extending this representation to 3D HD maps to account for elevation changes and complex terrain.
Summary updated