Scene‑aware and Fine‑grained Positioning Technology: Insights from Gaode Maps

At the 2019 Hangzhou Yunqi Conference, Gaode (Amap) presented a series of hot topics in travel‑related technologies, including visual and machine intelligence, route planning, scenario‑driven fine‑grained spatio‑temporal data applications, and the evolution of billion‑scale traffic architecture. This article is a concise transcript of the talk titled “Advancing Positioning Technology towards Scene‑aware and Fine‑grained Solutions” delivered by senior map technology expert Fang Xing.

The speaker highlighted that Gaode’s positioning service is not limited to the Gaode map app; it powers over 300,000 third‑party apps and serves billions of daily requests, handling roughly 10 11 positioning queries with millisecond‑level latency.

Challenges of Positioning

1. GPS alone provides ~10 m accuracy, which is insufficient for many scenarios such as distinguishing the side of a road while driving.

2. Indoor environments lack GPS signals, requiring alternative methods.

3. Balancing accuracy and cost demands large‑scale data mining, algorithmic improvements, and data‑quality enhancements.

Various techniques are employed: network‑based positioning (Wi‑Fi and cellular base‑station fingerprinting), inertial navigation, map‑matching, and emerging sensors such as visual, radar, and LiDAR.

Gaode’s Full‑scene Positioning Architecture

The system supports both mobile and in‑vehicle (car‑machine) scenarios. Mobile devices combine GPS with network positioning, while car‑machine scenarios add map‑matching and additional vehicle data.

Key problems include GPS blockage in underground parking, tunnels, or under high‑rise buildings, which affect ~60 % of failure cases. To mitigate this, Gaode integrates Wi‑Fi/base‑station data, map‑matching, and inertial navigation.

Fundamental Capabilities

Network positioning forms a data closed‑loop: each positioning request uploads the observed Wi‑Fi and base‑station lists, which are used both for immediate location estimation and for training two data products – (1) coarse Wi‑Fi base‑station locations and (2) detailed signal‑strength heatmaps. The more data collected, the better the training results, creating a virtuous cycle of accuracy improvement.

Algorithm Evolution

1. Clustering model : fast but low‑accuracy.

2. Grid‑based scoring : the space is divided into fine grids; each grid is scored using features such as historical point density, Wi‑Fi count, etc., improving 30 m accuracy coverage by 15 %.

3. Supervised machine‑learning : a neural network (including an RNN for historical trajectory) consumes base‑station/Wi‑Fi signal strengths, engineered features, and past positions to predict the most probable grid. This reduces large‑error cases by ~50 %.

To meet the requirement of handling billions of requests with <10 ms latency, three engineering optimizations are applied:

Hierarchical ranking : a coarse‑to‑fine cascade filters out unlikely locations early using large grids and few features.

Model simplification : depth and width of the neural network are reduced, and floating‑point precision is lowered for online inference.

Feature compression : an encoder compresses the high‑dimensional feature vector to a 2‑byte representation, cutting online feature storage by a factor of 10 (to <1 TB).

Scenario‑specific Accuracy Improvements

Indoor: data collection and Wi‑Fi‑to‑POI mapping raise indoor positioning accuracy by ~15 %. A CNN‑based image‑learning approach extracts building footprints from Wi‑Fi heatmaps, linking ~30 % of Wi‑Fi points to POIs.

Driving: a “soft+hard” fusion of mobile positioning, map‑matching, and sensor fusion (IMU, gyroscope) achieves >90 % of GPS points within 10 m, even in tunnels or parking lots.

Fusion positioning leverages low‑cost (<100 CNY) vehicle sensors, dynamic parameter calibration, and Hidden Markov Model (HMM) based map‑matching to correct drift and handle GPS outages.

Future Directions

• 5G: high‑frequency signals enable ranging, offering a new positioning modality complementary to Wi‑Fi and cellular.

• VSLAM and differential GPS: provide decimeter‑level accuracy for high‑precision maps required by lane‑level navigation and autonomous driving.

• Emerging ultra‑wideband, Bluetooth, and advanced Wi‑Fi standards will further enrich the sensor suite.

In summary, Gaode’s positioning platform combines massive data collection, iterative algorithmic upgrades, and system‑level optimizations to deliver scene‑aware, fine‑grained location services across mobile and automotive domains, while continuously exploring next‑generation technologies such as 5G‑based ranging and visual SLAM.