SGF-SLAM: Semantic Gaussian Filtering SLAM for Urban Road Environments

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous systems require robust environmental perception and mapping.
• Dynamic objects in road environments pose significant challenges for Simultaneous Localization and Mapping (SLAM) systems.
• Existing SLAM methods struggle with accuracy and stability due to dynamic elements.

Purpose of the Study:

• To propose an innovative SLAM framework for enhanced environmental perception and map construction in road environments.
• To suppress dynamic objects and improve mapping quality and system performance.
• To ensure accuracy and stability in mapping results for autonomous driving applications.

Main Methods:

• Developed a SLAM framework with a frontend tracking network (SGF-net) and a backend Semantic Gaussian Filter.
• Integrated feature point detection and semantic segmentation networks to filter dynamic objects.
• Introduced SGF-SLAM, a dynamic mapping framework utilizing multi-view geometry and semantic segmentation.
• Implemented a map update function using semantic information for occluded areas.

Main Results:

• SGF-net achieved over 23% improved inference speed compared to non-fused networks.
• The framework effectively suppresses dynamic objects (pedestrians, vehicles), restoring unobstructed road environments.
• Demonstrated significant enhancement in the reliability and adaptability of SLAM systems in road environments.

Conclusions:

• The proposed SGF-SLAM framework significantly enhances SLAM performance in dynamic road environments.
• The integration of semantic information and advanced filtering effectively handles dynamic objects.
• The system offers improved accuracy, stability, and adaptability for autonomous driving applications.