Enhancing SLAM algorithm with Top-K optimization and semantic descriptors.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Edge devices face computational challenges processing LiDAR data for deep learning.
• Efficient semantic mapping and loop closure detection are crucial for SLAM.

Purpose of the Study:

• To develop a SLAM algorithm that efficiently processes LiDAR point cloud data on edge devices.
• To reduce computational complexity and enhance processing speed for deep learning on LiDAR data.

Main Methods:

• Proposed a SLAM algorithm incorporating Top-K optimization for semantic descriptor and global semantic map generation.
• Extracted semantic information from LiDAR data to create 2D semantic descriptors.
• Implemented loop closure detection using geometric and semantic descriptor similarity, combined with front-end odometry.

Main Results:

• The Top-K strategy reduced average inference time by 10.7% and memory usage by 19.5%.
• The algorithm enhanced pose estimation accuracy and generated semantically rich global point cloud maps.
• More loop closures were matched without compromising accuracy.

Conclusions:

• The Top-K optimized SLAM algorithm significantly conserves computational resources for edge-device deep learning on LiDAR data.
• The approach effectively reduces computational load in practical applications while maintaining inference accuracy and efficiency.
• Improved semantic understanding and mapping capabilities for robots in their environment.