Comprehensive and Practical Vision System for Self-Driving Vehicle Lane-Level Localization

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving Systems

Background:

• Vehicle lane-level localization is crucial for autonomous driving.
• LIDAR systems are common but expensive and computationally intensive.
• There is a need for cost-effective and accurate localization solutions.

Purpose of the Study:

• To propose a low-cost stereovision system for accurate and consistent vehicle lane-level localization.
• To develop an integrated system combining advanced lane detection, road modeling, and stereo 3D reconstruction.
• To enhance localization consistency using a particle filter framework considering vehicle dynamics.

Main Methods:

• Integration of a novel lane line detection algorithm with existing detectors.
• Application of multiple road models for improved accuracy.
• Stereo 3D reconstruction for localization estimation.
• Utilizing a particle filter framework incorporating vehicle dynamics for consistent estimation.

Main Results:

• Lane line identification accuracy of 98.6%.
• Maximum estimation error for distance to lane lines: 16 cm (day), 26 cm (night).
• Maximum estimation error for moving direction: 0.06 rad (day), 0.12 rad (night).

Conclusions:

• The proposed stereovision system provides a practical and accurate solution for vehicle lane-level localization.
• It offers high performance across various visual conditions, outperforming limitations of current technologies.
• The system's accuracy and consistency make it suitable for implementation in autonomous vehicles.