Abstract
Autonomous vehicles typically rely on High-Definition (HD) maps for navigating complex urban intersections. However, these maps are static and fail in dynamic scenarios like road construction. This paper proposes a real-time, "anytime" method for estimating intersection layout and lane-level topology by observing the trajectories of surrounding vehicles. Using a two-step probabilistic approach involving Markov Chain Monte Carlo (MCMC) sampling and simulated annealing, the system achieves 99.9% topology accuracy and centimeter-level lane precision in under 150 milliseconds.
Introduction
The reliance on HD maps presents a significant vulnerability for autonomous driving systems. While highly detailed, these maps are "brittle"—they become obsolete during temporary road changes or accidents. Intersections are particularly problematic for mapless perception because visual cues like lane markings often fade or are occluded by other vehicles. This research introduces a novel paradigm: treating the collective trajectories of traffic participants as a high-level sensor for road structure. By observing where others drive, an autonomous vehicle can infer the underlying intersection geometry even when ground markings are invisible.
System Architecture
The proposed pipeline integrates data from multiple onboard sensors to track surrounding objects and estimate the road layout.
As shown in the architecture, detections from Radar, Lidar, and Camera are fused to track objects in a Bird's-Eye View (BEV). These tracked trajectories serve as the input for a two-stage estimation process:
- Topology Estimation: A coarse model identifying the intersection center, number of arms, and lane counts.
- Lane Course Estimation: A refinement stage that bends straight-line approximations into precise curves matching observed traffic flow.
Probabilistic Estimation and Refinement
The core of the system is a Markov Chain Monte Carlo (MCMC) sampling algorithm with simulated annealing. The algorithm generates various intersection hypotheses and scores them based on a likelihood function that measures how well the model explains observed vehicle positions and directions. This "anytime" property allows the system to provide an initial estimate almost immediately (approx. 36ms) and refine it as more computation or data becomes available.
The refinement process (illustrated above) demonstrates how the model evolves from a basic topological guess to a detailed lane-level representation. Unlike previous methods restricted to standard four-way junctions, this approach is "unrestricted," capable of modeling intersections with arbitrary numbers of arms and non-collinear roads.
Results and Discussion
The method was rigorously tested on both simulated data (1,000 intersections) and real-world data (14 intersections in Karlsruhe, Germany).
- Accuracy: The system correctly identified intersection topology with 99.9% accuracy in simulation.
- Precision: Lane course estimation achieved an average error of 15cm in simulation and 18-27cm in real-world urban scenarios.
- Efficiency: The entire pipeline completes within 113ms to 140ms, making it suitable for real-time integration into autonomous vehicle stacks.
Conclusion
By leveraging the aggregate intelligence of traffic flow, this research provides a robust alternative to HD maps for intersection navigation. While the system currently depends on the presence of other traffic, future work aims to integrate visual cues (e.g., lane markings, curbs) into the probabilistic framework to ensure reliability even in low-traffic conditions.
Summary updated