High‑Precision Positioning Techniques for Autonomous Driving in Complex Environments

Today we share high‑precision positioning technology for autonomous driving in complex environments.

Positioning/navigation is responsible for providing real‑time motion information of the vehicle, including position, speed, attitude, acceleration, and angular velocity.

Basic requirements for autonomous‑driving positioning systems are:

High precision: centimeter‑level accuracy.

High availability: stable operation in open‑world scenarios with complex conditions.

High reliability: positioning output feeds perception, planning and control, so any deviation can cause serious consequences.

Autonomous integrity detection: the system must warn the user and take measures when accurate output cannot be provided, keeping false‑alarm and missed‑alarm rates low.

To meet these requirements, the following methods are commonly used:

Inertial navigation (INS)

Global Navigation Satellite System (GNSS)

High‑definition map matching with online lidar point clouds

Odometer

Motion constraints based on vehicle dynamics

The Inertial Measurement Unit (IMU) is divided into two categories:

Fiber‑optic gyroscope (FOG) based IMU – high precision but costly, used in high‑accuracy mapping vehicles.

MEMS‑based IMU – compact, low‑cost, robust to environment, but with larger errors that require complex processing for autonomous‑driving test vehicles.

Raw IMU data are processed by a Strap‑down Inertial Navigation System (SINS) that includes:

Integrating gyro angular‑velocity to obtain attitude.

Transforming accelerometer specific force from the body frame to the navigation frame using the attitude.

Gravity, Earth‑rotation and other compensations.

Integrating acceleration to obtain velocity and position.

Because of integration, errors accumulate over time.

GNSS provides absolute positioning whose error does not grow with time. Modern GNSS has shifted from single‑frequency single‑system to multi‑frequency multi‑system (e.g., BeiDou, GLONASS, Galileo). A typical sky view in Beijing shows more than 35 satellites, greatly improving reliability and availability.

Precise positioning (RTK) based on carrier‑phase differential techniques offers centimeter‑level accuracy, dual‑antenna heading, and is widely used in intelligent driving tests, UAVs, precision agriculture, etc. However, it depends on base‑station infrastructure and network signals.

GNSS signals are vulnerable: urban canyons block low‑elevation satellites, increasing uncertainty; electromagnetic interference from vehicle electronics can degrade receiver performance.

HD‑map matching uses a pre‑built high‑definition map together with online lidar point clouds to achieve absolute, centimeter‑level positioning, but introduces dependence on the map.

Wheel sensors provide vehicle speed and odometry. External wheel sensors offer high resolution and accuracy but are complex and less reliable; internal sensors are compact and inexpensive but have lower accuracy and require extensive processing and online calibration.

Accurate calibration of the IMU axes relative to vehicle motion direction, online estimation of wheel parameters, and handling of wheel slip or bumps are necessary to keep errors bounded.

Vehicle motion constraints (e.g., detecting static state) can be used to improve robustness in extreme cases. These constraints are incorporated into a multi‑sensor fusion system.

After sensor data are obtained, the multi‑sensor fusion positioning pipeline consists of:

Data preprocessing: INS solution, GNSS quality control, lidar error compensation, wheel‑sensor calculations, and online estimation.

Map‑based matching using radar/lidar and HD maps.

Core modules: ZUPT/ZIHR/NHC – vehicle motion constraints INS Alignment – initial alignment of inertial navigation Integrated – combined fusion FDI – fault detection and isolation

Safety‑related module – integrity monitoring of the output.

The current navigation solution is based on a traditional Kalman filter whose objective is to minimize state variance. State variables include navigation parameter errors and sensor errors; the filter performs a prediction step followed by a measurement update.

Fault diagnosis and isolation employ statistical methods such as chi‑square testing and hardware redundancy. Deploying multiple GNSS/IMU units enables redundancy and improves reliability.

Typical scenarios illustrate how the system handles various challenges:

Open‑sky environment: GNSS/RTK and scan‑matching calibrate IMU/wheel‑sensor errors.

Bridge with weak longitudinal lidar features: GNSS, IMU, and wheel sensor detect scan‑match failure and keep positioning stable.

Lidar occlusion: the system must output reliable results without map matching.

Severe GNSS blockage (urban canyon, interference): sensor fusion must reject outliers and adapt parameters.

Under an overpass where satellite signals are unavailable: the system must operate independently of GNSS.

Guest Introduction

Cui Liuzheng, senior R&D engineer at Pony.ai, Ph.D. in Engineering from University of Chinese Academy of Sciences, leads multi‑source information fusion positioning research at Pony.ai. Previously worked at DJI on high‑precision positioning and GNSS/INS navigation.

Friend Recommendation

Join Pony.ai’s recruitment drive – follow “PonyAI小马智行” and submit your resume.

Event Recommendation

Join the Pony.ai Autonomous Driving Technology Salon – Shenzhen station. Scan the QR code to register.

Additional article recommendations are listed with links to Pony.ai’s infrastructure challenges, high‑definition maps, trajectory planning, data potential, and perception strategies.

About DataFun: a data‑intelligence platform that organizes offline deep‑tech salons and online content, aiming to share industrial experts’ practical experiences with the data‑science community.