How Visual‑Inertial Fusion Powers High‑Precision Maps for Autonomous Driving

Introduction

As navigation, driver assistance, and autonomous driving technologies evolve, maps must become increasingly detailed. Conventional road‑level maps fall short for intelligent transportation systems, prompting the use of visual‑inertial technology to create high‑precision maps.

Visual Sensors

Visual devices are classified by operation mode into monocular, stereo, and depth (RGB‑D) cameras. Monocular cameras are low‑cost but lack depth information, requiring motion for depth estimation. Stereo cameras provide depth via a known baseline but involve complex calibration and higher computational load. Depth cameras use infrared structured light for direct distance measurement, reducing computation but suffering from limited range, noise, narrow field of view, and sensitivity to sunlight, making outdoor deployment challenging. For large‑scale map production, monocular cameras are preferred due to their low cost and simple installation.

Inertial Sensors

Inertial navigation systems (INS) operate independently of external signals and are widely used in military, surveying, robotics, and autonomous driving. They measure acceleration in an inertial frame and integrate over time to obtain velocity, yaw, and position. INS units are categorized by drift rate: navigation‑grade, tactical‑grade, industrial‑grade, automotive‑grade, and consumer‑grade. Tactical‑grade INS is commonly selected for high‑precision positioning in autonomous driving and high‑precision map creation.

Modern INS variants include flexible, fiber‑optic, laser, and MEMS (Micro‑Electro‑Mechanical Systems). MEMS units are small, lightweight, low‑power, inexpensive, and robust, suitable for mid‑to‑low‑accuracy tactical applications. INS alone accumulates error; therefore, it is typically combined with GNSS (GPS, BeiDou) or other sensors to form a hybrid positioning system.

Fusion Framework and Key Technologies

The prevailing visual‑inertial fusion architecture consists of a front‑end that extracts sensor data and builds a model for state estimation, and a back‑end that optimizes the front‑end output to produce camera pose, attitude, and a global map. Gaode adopts a sliding‑window approach for local relative optimization, optionally initializing with pure visual Structure‑from‑Motion (SFM) aligned to INS when visual initialization fails. After local optimization, a global bundle adjustment refines the entire map.

Gaode High‑Precision Map Production

Map elements are divided into two categories: road signage (e.g., direction signs, traffic lights) and ground markings (e.g., lane dividers, arrows). Production involves three steps: (1) field data collection and trajectory solving to obtain visual‑inertial information; (2) automatic generation of map elements using the fusion pipeline; (3) manual web‑based editing to refine element accuracy before storing them in a database.

Examples of perception results and generated maps are shown below.

Future Outlook

Several companies (e.g., Moment, Kuandeng Technology, lvl5) are researching visual‑inertial high‑precision mapping, but current hardware costs limit achievable accuracy to about 10 cm. Future developments are expected to focus on multi‑source data fusion, where repeated captures from diverse devices are merged to improve both precision and map update latency.

Gaode’s extensive map data, automated production pipeline, and mature processes provide a solid foundation for next‑generation multi‑modal visual‑inertial mapping, further accelerating autonomous driving progress.