A Semantic-Enhanced Multi-Source Fusion Localization Method for GNSS-Degraded Environments.

Area of Science:

• Robotics
• Computer Vision
• Geomatics Engineering

Background:

• Global Navigation Satellite System (GNSS) signals are unreliable in urban environments due to signal blockage and multipath effects, leading to degraded positioning quality and accumulated errors in conventional navigation systems.
• Insufficient global constraints in existing integrated navigation systems exacerbate positioning inaccuracies, particularly in challenging GNSS-denied scenarios.

Purpose of the Study:

• To propose a novel multi-source integrated positioning method that incorporates semantic information to enhance navigation accuracy and robustness in GNSS-degraded urban environments.
• To develop a semantic target inversion model for transforming 2D image data into 3D position estimates using landmarks like traffic lights.
• To establish a noise covariance model for semantic factors to accurately weight their influence in factor graph optimization.

Main Methods:

• Utilized an object detection network to extract semantic landmark information (traffic lights) from images, including pixel coordinates and detection confidence.
• Developed a semantic target inversion model combining depth information, camera pose, and prior landmark coordinates to estimate 3D positions.
• Integrated semantic factors into a backend factor graph optimization framework, analyzing error influences to establish a noise covariance model for semantic factors.
• Validated the proposed method using an unmanned ground vehicle experimental platform under GNSS-degraded conditions.

Main Results:

• The proposed algorithm with semantic factors effectively suppressed accumulated positioning errors by providing supplementary global constraints.
• In Experiment 1, the maximum absolute trajectory error was reduced by 46.26% compared to the method without semantic factors.
• Experiment 2 demonstrated further robustness, reducing the maximum absolute pose error (APE) from 6.5432 m to 3.4778 m (approx. 46.85%) over a longer route with severe GNSS degradation.

Conclusions:

• The integration of semantic information, specifically from traffic lights, significantly improves the robustness and accuracy of multi-source fusion localization in GNSS-degraded environments.
• The developed semantic factor graph optimization method offers a viable solution for enhancing navigation system performance where traditional GNSS is unreliable.
• This research highlights the potential of leveraging semantic landmarks for reliable and accurate positioning in complex urban settings.