Abstract
As an emerging task that integrates perception and reasoning, topology reasoning in autonomous driving scenes has recently garnered widespread attention. However, existing work often emphasizes "perception over reasoning": they typically boost reasoning performance by enhancing the perception of lanes and directly adopt MLP to learn lane topology from lane query. This paradigm overlooks the geometric features intrinsic to the lanes themselves and are prone to being influenced by inherent endpoint shifts in lane detection. To tackle this issue, we propose an interpretable method for lane topology reasoning based on lane geometric distance and lane query similarity, named TopoLogic. This method mitigates the impact of endpoint shifts in geometric space, and introduces explicit similarity calculation in semantic space as a complement. By integrating results from both spaces, our methods provides more comprehensive information for lane topology. Ultimately, our approach significantly outperforms the existing state-of-the-art methods on the mainstream benchmark OpenLane-V2 (23.9 v.s. 10.9 in TOP$_{ll}$ and 44.1 v.s. 39.8 in OLS on subset_A). Additionally, our proposed geometric distance topology reasoning method can be incorporated into well-trained models without re-training, significantly boost the performance of lane topology reasoning.
Introduction
In the field of autonomous driving, perception is only the first step. A vehicle must not only detect lane lines but also understand their connectivity—the topology—to navigate complex environments like intersections. Current state-of-the-art models often prioritize perception over reasoning, relying on "black-box" Multi-Layer Perceptrons (MLPs) to infer topology from abstract lane queries. This approach is highly sensitive to "endpoint shifts," where minor perception errors in lane length lead to failures in connectivity reasoning. We propose TopoLogic, an interpretable pipeline that explicitly models lane topology through geometric and semantic relationships.
Methods
TopoLogic replaces opaque MLP-based reasoning with a transparent, two-pronged approach that integrates geometric distance and semantic similarity.
Geometric Distance Topology
Instead of implicit learning, TopoLogic explicitly calculates the distance between lane endpoints. To account for inevitable perception inaccuracies (endpoint shifts), we introduce a learnable mapping function ($f_{ours}$). This function is designed to be more "tolerant" than standard Gaussian curves, allowing the model to recognize connections even when predicted lane lines do not perfectly touch.
Lane Similarity Topology
To complement geometric data, TopoLogic projects lane queries into a high-dimensional semantic space. By calculating the dot-product similarity between these queries, the model identifies lanes that are semantically related, providing a robust fallback when geometric detection is noisy or endpoints are significantly shifted.
Topology-Enhanced Learning
The geometric and similarity matrices are fused into a final topology graph. This graph is processed by a Graph Neural Network (GNN), enabling feature propagation between connected lanes. This interaction allows the model to refine its scene understanding based on the established network structure.
Key Results
We evaluated TopoLogic on the OpenLane-V2 benchmark, the most comprehensive dataset for lane topology. Our method sets a new state-of-the-art across multiple metrics and subsets.
On Subset_A, TopoLogic achieved a TOP$_{ll}$ score of 23.9, more than doubling the performance of the previous SOTA, TopoNet (10.9). The overall score (OLS) also saw a significant improvement from 39.8 to 44.1. Notably, the geometric distance module can be applied as a "plug-and-play" post-processing step to existing models. For instance, adding it to a pre-trained TopoNet model without any re-training boosted its TOP$_{ll}$ from 10.9 to 22.3, proving the fundamental superiority of explicit geometric logic over MLPs.
Discussion and Conclusion
TopoLogic demonstrates that explicit, interpretable reasoning is a powerful alternative to purely data-driven MLP approaches in autonomous driving. By making the reasoning logic robust to perception errors, we achieve substantial performance gains and provide developers with a clear "Distance Matrix" to debug connectivity decisions.
One current limitation is that while improved topology reasoning enhances scene understanding, it has a limited impact on the initial detection of lane lines. Future work will explore "closing the loop," using topological constraints to provide feedback to the perception backbone, potentially improving the detection of obscured or faded lanes based on the road's inherent logic.
Summary updated