Efficient Underground Tunnel Place Recognition Algorithm Based on Farthest Point Subsampling and Dual-Attention...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous navigation in GPS-denied environments like underground tunnels is critical.
• Existing place recognition algorithms struggle with sparse and noisy point cloud data common in tunnels.
• Ensuring high recognition accuracy and robustness is a significant challenge.

Purpose of the Study:

• To propose an efficient point cloud place recognition algorithm, Dual-Attention Transformer Network (DAT-Net), for underground tunnels.
• To enhance the utilization of effective features in limited underground tunnel data.
• To improve the accuracy and robustness of autonomous place recognition systems.

Main Methods:

• Implemented a farthest point downsampling module to reduce point cloud size and retain essential shape information.
• Developed a dual-attention Transformer module employing multi-head self-attention for local information exchange.
• Integrated local descriptors with global descriptors using a feature fusion layer for robust global feature representation.

Main Results:

• The proposed DAT-Net achieved an average F1 score of 0.841 on the SubT-Tunnel dataset.
• Demonstrated superior recognition accuracy compared to existing state-of-the-art algorithms.
• Showcased enhanced robustness in place recognition tests within challenging tunnel environments.

Conclusions:

• DAT-Net offers an effective solution for autonomous place recognition in GPS-inaccessible underground tunnels.
• The algorithm's dual-attention mechanism and feature fusion enhance the processing of sparse point cloud data.
• The findings indicate significant advancements in robust and accurate robotic navigation within subterranean environments.