SGC-VSLAM: A Semantic and Geometric Constraints VSLAM for Dynamic Indoor Environments

Area of Science:

• Robotics and Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Visual Simultaneous Localization and Mapping (VSLAM) is crucial for autonomous mobile robots.
• Existing VSLAM methods often assume scene rigidity, limiting performance in dynamic environments.

Purpose of the Study:

• To develop a VSLAM system robust to dynamic objects.
• To improve localization accuracy and enable static map construction in real-world scenarios.

Main Methods:

• Introduced Semantic and Geometric Constraints VSLAM (SGC-VSLAM) building upon ORB-SLAM2.
• Integrated dynamic feature detection using semantic (YOLO v3) and geometric constraints.
• Implemented an improved quadtree-based feature extraction and a static point cloud map construction module.

Main Results:

• SGC-VSLAM effectively filters dynamic features using semantic and geometric constraints.
• The system demonstrated improved positioning accuracy in high-dynamic scenarios on the TUM RGB-D dataset.
• Successfully constructed static point cloud maps of the environment.

Conclusions:

• SGC-VSLAM enhances VSLAM robustness and accuracy in dynamic environments.
• The ability to build static maps has significant long-term value for autonomous mobile robots.