Visual Intelligence Connecting the Real World – Amap’s Mapping and AR Navigation Technologies

At the 2019 Hangzhou Yunqi Conference, the Amap (Gaode) technology team presented a series of hot topics in travel technology, including visual and machine intelligence, route planning, fine‑grained positioning, spatio‑temporal data applications, and the evolution of billion‑scale traffic architecture.

Chief Scientist Ren Xiaofeng delivered a talk titled “Visual Intelligence Connecting the Real World”. The core mission of Amap is expressed as “Connecting the real world to make travel better”.

Why vision is the bridge

Human perception obtains about 80% of information through vision, and the brain devotes 30‑60% of its processing power to visual perception. From a machine perspective, vision is a universal sensing modality that provides rich, long‑range, real‑time information.

In map production, visual data is the primary source. Over 1 billion daily active users rely on Amap, which not only provides navigation but also services such as shared mobility, smart bus, smart scenic spots, cycling, walking, and long‑distance travel.

Visual technology in map making

Data collection is mainly performed by cameras. The sheer scale of national road networks (millions of kilometers) makes manual processing infeasible; automated visual recognition is essential, with human correction used only when algorithms fall short.

Typical map‑making tasks are divided into road‑related work and POI (point‑of‑interest) sign recognition. Road‑sign detection must locate every sign, classify its type, and read its content.

Challenges include low‑cost data acquisition (image quality, distortion, glare, occlusion), camera intrinsic/extrinsic parameters, and the need for geometric correction. Multi‑source data matching and distortion correction significantly improve downstream algorithm performance.

Image quality issues are addressed with enhancement techniques that detect blur, improve clarity, and boost recognition accuracy.

Detecting tiny targets such as traffic cameras (electronic eyes) requires zoom‑in strategies and attention mechanisms that let the model focus on the most informative regions while ignoring irrelevant background.

Beyond road signs, POI detection must differentiate between true POIs and other signage (advertisements, banners, etc.) and often requires multi‑task models that combine text detection, recognition, and 3‑D layout understanding.

Because no single algorithm can solve all problems, Amap employs ensembles of models—e.g., separate detectors for text, geometry, and semantics—integrated on a large accumulated dataset.

Visual technology in navigation

Amap’s AR navigation, released in April, overlays AI‑generated cues (lane guidance, turn arrows, speed limits, collision warnings) directly onto the video feed captured by a low‑cost camera. The system aims to provide “what‑you‑see‑is‑what‑you‑get” guidance with computational demand roughly one‑fifth of a typical smartphone.

To achieve this, the team uses model compression, small‑model training, joint detection‑tracking, multi‑target unified models, and tight integration with traditional GPS navigation.

The overall goal is to connect the real world more accurately and affordably, enabling higher‑quality map production and precise positioning for everyday users.