A Lightweight Visual Simultaneous Localization and Mapping Method with a High Precision in Dynamic Scenes.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Traditional visual SLAM (VSLAM) systems struggle with accuracy in dynamic environments due to static assumptions.
• Achieving high accuracy in dynamic scenes often compromises real-time performance, creating a critical need for balanced solutions.

Purpose of the Study:

• To propose a novel VSLAM system that effectively balances positioning accuracy and computational complexity.
• To improve the overall performance of VSLAM systems in highly dynamic environments.

Main Methods:

• Implemented an improved lightweight target detection network with an attention mechanism to identify and exclude dynamic feature points.
• Utilized the Ghostnet module as the backbone for YOLOv5s to reduce model parameters and enhance inference speed.
• Focused on matching only static feature points for improved frame accuracy.

Main Results:

• The proposed VSLAM system demonstrated an 84.04% improvement in pose estimation accuracy compared to ORB-SLAM3 on the TUM dynamic dataset.
• Significantly improved positioning accuracy compared to DS-SLAM and DVO SLAM.
• Achieved superior real-time performance while maintaining comparable accuracy to other deep learning-based VSLAM algorithms.

Conclusions:

• The developed VSLAM system offers a robust solution for accurate and real-time localization in dynamic environments.
• The integration of lightweight networks and attention mechanisms provides an effective strategy for overcoming the limitations of traditional VSLAM.
• This research contributes to advancing VSLAM technology for applications requiring high precision and efficiency in complex, dynamic scenes.