Driving by the Rules: Integrating Lane-Level Traffic Regulations into Online HD Maps

Gaode Map, a leading provider of travel and location services, highlights the importance of lane‑level driving rules for autonomous driving. While most research focuses on constructing vector maps, the integration of lane‑level traffic regulations is often overlooked.

The authors argue that compliance with traffic rules is essential for autonomous systems, and that lane‑level rules are typically embedded in high‑definition (HD) maps. However, the low update frequency and high cost of HD maps limit online perception‑based mapping. Existing online mapping methods mainly capture road structures and ignore the semantic traffic rules, forcing systems to rely on offline maps.

To address this gap, Gaode Map and Xi'an Jiaotong University propose the "Driving by the Rules" concept, introducing a new task: integrating traffic‑sign‑based lane‑level regulations into online‑constructed HD maps, and they release a benchmark called MapDR for researchers.

The MapDR dataset contains over 10,000 real driving scenes from Beijing, Shanghai, and Guangzhou, with more than 18,000 structured lane‑level traffic rules. Each scene provides raw data (continuous front‑view images, sign poses, vectorized maps, camera intrinsics and poses) and annotations (rules with associated lane centerlines and sign regions).

The proposed task consists of two sub‑tasks: (1) extracting lane‑level traffic rules from traffic signs, and (2) associating each rule with the corresponding lane centerline. The evaluation treats the overall problem as a bipartite graph matching, measuring precision and recall for rule nodes, edge predictions, and the minimal sub‑graph composed of a rule, a centerline, and an edge, with F1 score as the final metric.

Two baseline approaches are presented:

Modular Approach (VLE‑MEE) : a three‑stage pipeline—Grouping (fusing sign images and OCR to group lane‑level regions), Understanding (classifying each group to predict rule key‑value pairs), and Association (encoding vector maps and fusing with rule features to predict lane‑rule links via a binary classifier).

End‑to‑End Approach (RuleVLM) : builds on the multimodal large language model Qwen‑VL‑Chat 7B, exploring LoRA fine‑tuning with different prompts (TextPrompt, VisualPrompt) and a best configuration that feeds full front‑view and sign images together with vector features extracted by MEE.

Experiments on the MapDR dataset show that heuristic OCR‑plus‑nearest‑lane methods struggle with complex signs, while VLE‑MEE and RuleVLM provide effective baselines for both modular and end‑to‑end modeling.

In conclusion, the paper contributes (1) the new task of integrating traffic regulations into online HD maps, (2) the MapDR dataset and evaluation metrics, and (3) two baseline methods (VLE‑MEE and RuleVLM) to advance research in this area.