How Amap Engineers Precise, Real‑Time Intersection Data with AI and Optimization

01 How to Express Intersections

Guiding users safely and efficiently through complex intersections is a core challenge for navigation software. For city traffic managers, improving intersection throughput eases congestion; for drivers, following lane guidance enables smooth passage; for assisted driving, accurate lane‑level data enhances safety.

Amap's data team addresses these challenges by simplifying intersection model representation, automating topology generation, and using multimodal large‑model updates, providing a solid data foundation for routing and guidance engines.

The left image shows high‑fidelity lane‑level data useful for lane‑level navigation and advanced driver assistance; the right image shows an abstracted road skeleton that stores topology, geometry, and attributes with minimal memory for routing and traffic calculations.

Intersection Model Design Challenges

1. Diversity: Real‑world intersections vary widely (cross, T‑shaped, offset, irregular, roundabouts, etc.).

2. Complexity: The abstract skeleton influences routing, guidance, traffic estimation, and map rendering, acting as the map’s neural core.

3. Uncertainty: Multiple geometric representations exist, making it hard to determine the optimal design solely by experience.

Amap’s Solution

1. Intersection Types: By analyzing surrounding road structure, environment, entry/exit counts, and connection angles, Amap categorizes over 50 real‑world intersection types, resulting in more than 500 combinatorial skeletons. These are evaluated against application requirements to select the best geometric connection.

2. Model Design: The geometric connection problem is reduced to a p‑median optimization, treating intersection nodes as service stations and stop‑line nodes as demand points, minimizing the total weighted distance.

3. Data Specification: Geometric algorithms optimize the combination of topology, geometry, and attributes to minimize storage size.

Precise Intersection Production

Amap uses professional surveying equipment combined with AI‑assisted workflows to ensure high‑quality data.

1. Professional Collection: Regular autonomous surveys, massive crowdsourced data, and sophisticated dispatch algorithms quickly gather high‑precision road information.

2. Data Production: Multimodal large‑model algorithms automatically recognize, generate, and fuse elements such as road boundaries, lane markings, stop lines, crosswalks, and channelizing strips, achieving low‑cost, high‑automation production.

Real‑Time Intersection Updates

Rapid growth in vehicle numbers and road changes demands frequent updates.

Challenges:

1. Data sparsity across regions and time periods. 2. Noise from random user behavior. 3. Complex temporal patterns due to diverse temporary constructions.

Technical Solutions:

1. Real‑time change detection using VLM multimodal perception to capture alterations in road shape, lane count, and arrow types at daily or hourly granularity.

2. Low‑cost updates by invoking the data‑production algorithms described earlier.

3. High‑quality model training through a curated sample‑building system and automated labeling to address long‑tail and difficult cases.

Navigation AI Agent

Accurate data enables Amap’s self‑developed AI navigation agent, transforming navigation from a tool into a smart travel companion. Using a Planner‑Executor architecture, it closes the perception‑planning‑execution‑expression loop, employs TrafficVLM for precise traffic capture, and AmapVoice for expressive, empathetic guidance.