How High‑Precision Maps Power Autonomous Driving: Inside Amap’s AI and Cloud Strategies

Background

Amap (Gaode) presented at the 2020 Cloud Ridge Conference a session titled “Smart Mobility,” focusing on the development of a next‑generation travel service platform built on DT+AI and a fully cloud‑native architecture. The talk highlighted high‑precision maps, algorithms, AR navigation, lane‑level technology, and their role in autonomous driving.

Why High‑Precision Maps Matter

High‑precision maps are essential for autonomous vehicles, providing centimeter‑level positioning, static object modeling, and path/behavior planning. Amap aims for a relative accuracy of 10 cm over 100 m, requiring frequent updates (from yearly to daily) to maintain map freshness.

Key Challenges

Accuracy – achieving 10 cm relative precision.

Scale – handling massive city‑wide datasets.

Cost – expensive LiDAR and GNSS equipment.

Timeliness – reducing update cycles.

Algorithmic Solutions

The workflow is divided into three algorithmic pillars:

Data precision and alignment.

Recognition and production automation.

Change detection and map updating.

Data Precision & Alignment

Front‑end processing relies on point‑cloud matching (ICP, GICP) and fast semantic segmentation of point clouds and lane lines. Back‑end performs large‑scale optimization to jointly adjust trajectories across an entire city, using sparse spline‑based optimization for hundred‑fold speed‑ups.

Recognition & Production Automation

After alignment, the pipeline extracts map elements (lane lines, guardrails, poles, traffic signs, etc.) from point clouds and images. Techniques include multi‑level random aggregation networks for point‑cloud semantic segmentation, image panorama segmentation, point‑cloud/image fusion, traditional fitting algorithms, and vectorization with deep features.

Change Detection & Updating

Amap employs two update pipelines:

Laser‑based: low‑cost LiDAR with tightly coupled SLAM/LIO, real‑time semantic segmentation, multi‑layer localization maps, in‑loop relocalization, covariance modeling, and global pose optimization.

Vision‑based: consumer‑grade cameras and IMUs, using visual SLAM/VIO, feature‑semantic layers, VIO‑based relocalization, and global pose optimization.

Both pipelines achieve industry‑leading precision (≈15 cm) and high recall (>98 %).

Results and Outlook

The described methods enable high‑accuracy map creation, large‑scale alignment, and rapid, low‑cost updates, positioning Amap’s high‑precision mapping as a core component of future autonomous driving solutions. Ongoing work aims to further improve accuracy and update frequency.