How Mobile and Car Navigation Achieve Precise Positioning: Sensor Fusion, Map Matching, and High‑Precision Evolution

1. Navigation Positioning Framework

The core goal of navigation positioning is to provide continuous, reliable location data for navigation services, such as determining the current road, detecting route deviation, and measuring distance to the next intersection.

Typical input signals include GPS (5‑10 m accuracy), inertial sensors (gyroscope, accelerometer), magnetometer, barometer, and, for vehicle units, CAN‑bus data like vehicle speed and steering angle. These signals are fused using attitude integration and dead‑reckoning algorithms, then matched to map data to associate the device’s position with road geometry.

2. Mobile Navigation Positioning

2.1 Attitude Fusion (AHRS)

Six‑axis inertial sensors (gyroscope + accelerometer) are fused using AHRS algorithms such as complementary or Kalman filters. Gyroscope integration yields angular change, while accelerometer data (including gravity) provides tilt angles. For nine‑axis sensors, magnetometer data is incorporated into the same fusion framework.

2.2 Map Matching

Traditional map‑matching relies on geometric criteria (distance, heading) near the GPS point, which becomes unstable under multi‑meter GPS errors. A robust solution combines multiple sensor inputs and uses a hidden Markov model (HMM) as the core matcher, supplemented by scenario‑specific strategies.

In the HMM formulation, the hidden state z_n represents the matched road segment, while the observation x_n is the GPS measurement. The Viterbi algorithm computes the most probable road sequence.

Emission probabilities are modeled as:

Position: a normal distribution where probability increases with proximity to a road.

Heading: a Von Mises distribution weighted by vehicle speed.

Transition probabilities consider road curvature and speed, using Von Mises for turning angle and exponential distribution for travel distance consistency.

The HMM‑based matcher implemented in the Gaode mobile app significantly improves matching accuracy and stability compared with earlier rule‑based methods.

3. Vehicle (Car‑Headunit) Navigation Positioning

3.1 Positioning Scheme

Vehicle navigation must handle complex scenarios such as tunnels, underground parking, and urban canyons. Multi‑sensor fusion is essential: when GPS degrades, vehicle speed pulses and inertial data provide dead‑reckoning, while map feedback corrects accumulated drift.

Different sensor configurations lead to distinct schemes, ranging from pure GNSS (lowest performance) to front‑end fusion (IMU + speed) and full back‑end fusion that integrates GNSS, IMU, CAN‑bus data, and map information for the best accuracy.

3.2 Sensor Fusion Technique

The back‑end fusion framework follows a Kalman‑filter architecture, with state‑transition and observation models built from vehicle dynamics and sensor noise characteristics. The filter simultaneously estimates navigation states and sensor biases (e.g., gyro bias, speed‑sensor scale).

Experimental results show that, even with low‑cost sensors, the fused solution can match or exceed the performance of expensive professional inertial systems.

4. High‑Precision Positioning Evolution

Traditional navigation meets road‑level accuracy (≈10 m). Emerging assisted‑driving and autonomous‑driving use cases demand lane‑level (meter to sub‑meter) and eventually centimeter‑level precision.

Two upgrade paths exist:

Upgrade existing sensors (e.g., RTK‑GPS, high‑grade MEMS IMU) while keeping the algorithmic framework.

Introduce new sensors such as LiDAR, millimeter‑wave radar, and cameras, requiring new fusion algorithms.

Gaode’s collaboration with Qianxun produced a RTK‑based solution “ZhiTu” achieving <10 cm accuracy without additional sensors or map data.

Relative‑positioning approaches include LiDAR‑based SLAM (mature but costly) and vision‑based SLAM, which benefits from rapid advances in visual algorithms and edge‑computing hardware. Gaode plans three high‑precision services:

System‑level positioning for L3 autonomous driving, fusing external visual semantics, HD‑maps, GPS/RTK, and IMU.

Lane‑level navigation using proprietary visual algorithms and cloud‑based image localization.

Crowdsourced high‑precision data collection using low‑cost visual+RTK+IMU hardware and vSLAM.

Conclusion Traditional navigation relies on 10 m GPS accuracy and adapts sensor‑fusion, behavior‑recognition, and map‑matching techniques for mobile and vehicle platforms. Future autonomous‑driving scenarios push accuracy toward lane‑level and centimeter‑level, demanding continuous algorithmic and sensor innovations.