LVID-SLAM: A Lightweight Visual-Inertial SLAM for Dynamic Scenes Based on Semantic Information

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for robot navigation but faces challenges in dynamic environments.
• Existing deep learning SLAM methods are either fast but inaccurate or accurate but computationally expensive.
• Current maps lack semantic information, limiting robot environmental understanding and task performance.

Purpose of the Study:

• To develop a lightweight visual-inertial SLAM system capable of operating effectively in dynamic environments.
• To enhance robot environmental understanding by integrating semantic information into the mapping process.
• To improve pose accuracy and robustness compared to existing SLAM frameworks.

Main Methods:

• The system, LVID-SLAM, is built upon the ORB-SLAM3 framework, incorporating a new thread for object detection.
• It tightly couples semantic object detection with geometric information to filter dynamic features.
• Inertial Measurement Unit (IMU) data is utilized to aid feature extraction and recover from visual tracking loss, constructing a dense octree-based semantic map.

Main Results:

• LVID-SLAM demonstrated excellent pose accuracy and robustness in highly dynamic scenes.
• Achieved an average Absolute Trajectory Error (ATE) reduction of over 80% compared to ORB-SLAM3 on the TUM dataset.
• Outperformed other methods in dynamic conditions, providing both real-time performance and robustness.

Conclusions:

• The proposed LVID-SLAM system offers a robust and efficient solution for visual-inertial SLAM in dynamic environments.
• Integrating semantic information and IMU data significantly enhances SLAM performance.
• The system provides a foundation for robots with improved environmental awareness and task execution capabilities.