IS-CAT: Intensity-Spatial Cross-Attention Transformer for LiDAR-Based Place Recognition

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• LiDAR place recognition is vital for autonomous navigation and Simultaneous Localization and Mapping (SLAM).
• LiDAR offers robustness in challenging environments where camera-based methods falter due to weather or lighting variations.

Purpose of the Study:

• To introduce a novel LiDAR-based approach for enhanced place recognition.
• To explore the synergy between spatial and intensity data in LiDAR for global descriptor generation.

Main Methods:

• Developed the intensity and spatial cross-attention transformer (IS-CAT) model.
• Utilized a cross-attention to concatenation mechanism to integrate multi-layered LiDAR projections.
• Fused spatial and intensity LiDAR data for comprehensive place representation.

Main Results:

• IS-CAT demonstrated superior performance in place recognition tasks on the NCLT and Sejong indoor-5F datasets.
• The model showed successful application within a 3D LiDAR SLAM system.
• Achieved enhanced performance in both indoor and outdoor environments.

Conclusions:

• The proposed IS-CAT method effectively fuses spatial and intensity LiDAR data for advanced place recognition.
• This approach offers practical effectiveness and significant advancements for autonomous navigation systems.
• The findings underscore the value of integrating multi-modal LiDAR data for robust localization.