YPD-SLAM: A Real-Time VSLAM System for Handling Dynamic Indoor Environments.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for autonomous systems but is often disrupted by dynamic elements in real-world environments.
• Existing Visual SLAM (VSLAM) algorithms struggle with accuracy and speed when faced with dynamic objects, limiting their practical application.

Purpose of the Study:

• To develop a real-time VSLAM algorithm robust to dynamic indoor environments.
• To enhance localization accuracy and processing speed compared to existing SLAM methods.

Main Methods:

• Utilized a lightweight YoloFastestV2 deep learning model with NCNN and MNN for semantic image information.
• Implemented epipolar constraints and object dynamic properties to remove dynamic feature points.
• Introduced Cylinder and Plane Extraction (CAPE) for planar detection and integrated planar constraints into SLAM's nonlinear optimization.

Main Results:

• Achieved significant average improvements in localization accuracy: 91.95% over ORB-SLAM2, 27.21% over DS-SLAM, and 30.30% over RDMO-SLAM on dynamic sequences.
• Reduced single-frame tracking time to 42.68 ms, outperforming DS-SLAM, RDMO-SLAM, and RDS-SLAM by 14.6-34.33%.

Conclusions:

• The proposed VSLAM algorithm demonstrates superior real-time performance and localization accuracy in dynamic indoor environments.
• The system's efficiency and robustness make it easily deployable on various platforms for enhanced autonomous navigation.