Unlocking Indoor Positioning: From Geometry to Wi‑Fi Fingerprinting

Introduction

Map software has become essential for daily travel, but satellite positioning cannot reliably work indoors; therefore, indoor positioning is a key research focus for Baidu Maps.

Geometric Positioning Algorithms

TOA (Time of Arrival)

TOA measures the travel time of electromagnetic waves from a device to three reference points (t1, t2, t3). Using the speed of light, distances r1, r2, r3 are computed and the target location is solved via planar geometry.

TDOA (Time Difference of Arrival)

TDOA uses the difference in arrival times to two reference points, placing the target on a hyperbola. This eliminates the need to measure the absolute start time, reducing timing error.

AOA (Angle of Arrival)

AOA measures the incident angle of a signal from two reference points, allowing the target position to be determined with fewer reference nodes, though it requires directional antenna arrays.

RSSI (Received Signal Strength Indicator)

RSSI uses a signal‑strength‑to‑distance attenuation model. By measuring signal strength from three reference points, distances are estimated and the target location is solved geometrically.

Professional‑Equipment Based Positioning

Based on the physical nature of signals, indoor positioning methods can be divided into electromagnetic wave‑based (UWB, Wi‑Fi, Bluetooth, ZigBee, infrared, LED) and ultrasonic (mechanical wave) techniques, each with distinct accuracy, cost, and deployment characteristics.

Wi‑Fi for Indoor Positioning

Because Wi‑Fi access points are already densely deployed in malls and public spaces, they can be leveraged for indoor positioning. The basic idea is to map scanned Wi‑Fi signal strengths (RSSI) to coordinates via a function F.

Fingerprint Method

Since modeling the exact Wi‑Fi signal distribution is difficult, the fingerprint approach directly samples signal data at many locations and stores them in a database. During positioning, a device’s current Wi‑Fi scan is matched against stored fingerprints using a K‑Nearest‑Neighbour (KNN) style weighted calculation.

Local Model + Fingerprint

When fingerprint data are sparse, a local propagation model can generate virtual fingerprints for unmeasured areas, improving coverage while maintaining accuracy.

Position Regression Analysis

The relationship between scanned AP signals and coordinates can be treated as a regression problem. Machine‑learning models (e.g., decision trees) are trained to predict x and y from the signal feature vector.

Geomagnetic Positioning

Earth’s magnetic field, perturbed by building structures, provides a stable fingerprint. By matching measured magnetic sequences to a pre‑collected database (using spatial indexing and particle filtering), high‑accuracy indoor positioning can be achieved.

Conclusion

Indoor positioning combines signal‑processing expertise with machine‑learning knowledge. Successful deployment requires extensive technical accumulation and experience. Baidu aims to provide a world‑class indoor positioning service that delivers high‑precision location experiences to users.