How Visual AI Powers Real-World Mapping and AR Navigation at Amap

Visual AI: Bridging the Real World

Human beings obtain over 80% of information through vision, and the brain devotes 30‑60% of its resources to visual perception; similarly, machines treat vision as a universal sensing modality because it provides rich, long‑range, real‑time data.

Visual Technology at Amap: Map Production

Amap serves more than 100 million daily active users and over 400 million monthly active users, offering navigation, ride‑hailing, smart bus, scenic‑spot guides, cycling, walking, and long‑distance travel services. Building a reliable map requires massive visual data collection across millions of kilometers of roads. Cameras capture images that are later processed by algorithms to detect road signs, POI (point‑of‑interest) signs, and other assets. The workflow combines automatic recognition with manual correction to create a map database that supports downstream services.

Typical map‑making tasks are divided into two categories: road‑related data and POI sign recognition. Road‑sign detection involves locating each sign, classifying its type, and extracting its content. Real‑world image acquisition faces challenges such as distortion, glare, occlusion, low resolution, and adverse weather, which demand robust preprocessing, distortion correction, and multi‑source data matching.

Beyond standard signs, Amap must handle tiny targets such as electronic eyes, which require attention‑based mechanisms and region‑wise scaling to improve detection accuracy while avoiding overwhelming background information.

Visual Technology at Amap: Navigation

Traditional turn‑by‑turn guidance can be confusing in complex intersections. Amap aims to provide “what‑you‑see‑is‑what‑you‑get” navigation by overlaying AI‑generated cues directly onto live video streams from vehicle‑mounted cameras. The AR navigation product, launched in April, offers lane‑keeping prompts, turn arrows, speed limits, collision warnings, and pedestrian alerts.

Achieving this on low‑cost hardware is a major challenge. Unlike autonomous‑driving stacks that use high‑power dedicated chips, Amap’s solution must run on computational budgets comparable to a typical smartphone, requiring model compression, small‑model training, detection‑tracking fusion, multi‑task networks, and tight integration with GPS.

These visual AI techniques improve map‑making efficiency, navigation accuracy, and overall user experience, aligning with Amap’s mission to “connect the real world and make travel better.”