RGB-D Visual SLAM Based on Yolov4-Tiny in Indoor Dynamic Environment

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Simultaneous Localization and Mapping (SLAM) systems face challenges in dynamic indoor environments due to moving obstacles.
• Reduced position estimation accuracy and visual odometer stability are common issues in such scenarios.

Purpose of the Study:

• To propose a robust visual SLAM algorithm for dynamic indoor environments.
• To enhance camera position estimation accuracy and system stability by addressing the impact of moving objects.

Main Methods:

• A visual SLAM algorithm integrating the Yolov4-Tiny network for object detection and semantic information extraction.
• A dynamic feature point elimination strategy based on traditional ORBSLAM, utilizing epipolar geometry and LK optical flow for dynamic object detection.
• Filtering out dynamic feature points in the tracking thread to rely solely on static points for camera pose estimation.

Main Results:

• The proposed algorithm demonstrated a 93.35% improvement in camera position estimation accuracy compared to ORB-SLAM2 in highly dynamic environments.
• Achieved real-time performance with an average processing time of 21.49 ms per image frame in the tracking thread.
• Validated on the TUM dataset, confirming its effectiveness in challenging indoor conditions.

Conclusions:

• The developed visual SLAM method effectively handles dynamic elements in indoor environments.
• The integration of object detection and dynamic feature point elimination significantly enhances localization accuracy and stability.
• The algorithm offers a practical solution for real-time robotic navigation in complex, changing indoor settings.