Abstract
Localization is a fundamental challenge for autonomous vehicles. While many approaches focus on final pose estimation, the underlying processes of data representation and data association are critical for accuracy. This paper introduces DA-LMR (Delta-Angle Lane Marking Representation), a 3D representation that encodes local curvature into 2D lane markings, and DC-SAC (Distance-Compatible Sample Consensus), a heuristic data association method. By incorporating geometric context and distance constraints, these methods significantly improve localization robustness in noisy environments, achieving high precision and recall even with significant sensor error.
Introduction
For a self-driving vehicle to localize itself, it must match real-time sensor detections (e.g., lane markings from cameras or LiDAR) with pre-existing map landmarks. This "data association" step is often hindered by two factors: information loss in simple 2D point representations and the computational complexity of matching in the presence of noise and outliers. Standard point-based matching often fails to distinguish between similar-looking straight segments, leading to incorrect pose estimates.
Methodology
DA-LMR: Delta-Angle Lane Marking Representation
The core innovation of this work is the addition of a third dimension to traditional 2D lane marking points. Instead of a simple $(x, y)$ coordinate, each point is represented as $(x, y, \Delta\alpha \cdot w)$.
As shown in the figure above, for any point $p_i$ on a polyline, the Delta-Angle ($\Delta\alpha_i$) is calculated as the change in orientation between the segment entering the point and the segment leaving it. This value effectively encodes the local curvature. By weighting this angle ($w$), the representation provides a unique geometric signature for each point, making curved segments highly distinguishable from straight ones and ensuring the representation is invariant to the vehicle's driving direction.
DC-SAC: Distance-Compatible Sample Consensus
To improve the efficiency and reliability of the matching process, the authors propose DC-SAC. This method enhances the standard RANSAC algorithm by introducing a Distance Compatibility heuristic. Rather than testing random pairs of points, DC-SAC only considers hypotheses where the Euclidean distance between two sensor detections is consistent with the distance between their corresponding map landmarks. This constraint prunes the search space, allowing for faster convergence and better outlier rejection.
Experimental Results
The proposed methods were evaluated using synthetic data derived from aerial imagery of Karlsruhe, Germany, and real-world data from the BerthaOne experimental vehicle. Testing involved varying levels of Gaussian noise (up to $\sigma = 0.5$m) and the addition of 10% outliers.
As illustrated in the performance plots, the DA-LMR representation consistently outperforms standard Point Representation (PR) and Line Segment Representation (LSR). Even at high noise levels (0.5m standard deviation), the combination of DA-LMR and DC-SAC maintains superior F1-scores, precision, and recall compared to state-of-the-art alternatives. Quantitatively, the system reached up to 98.1% precision and 99.7% recall in challenging scenarios.
Conclusion
DA-LMR and DC-SAC provide a robust framework for the data association sub-problem in autonomous localization. By treating local geometry as a first-class feature and enforcing spatial consistency during matching, these methods overcome the limitations of traditional point-based systems. Future work will focus on extending this representation to 3D environments and automating the tuning of the curvature weight parameter.
Summary updated