Enhanced Simultaneous Localization and Mapping Algorithm Based on Deep Learning for Highly Dynamic Environment.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual Simultaneous Localization and Mapping (SLAM) is crucial for autonomous navigation.
• Traditional SLAM methods face accuracy limitations in dynamic environments due to unpredictable changes.
• Asymmetric information acquisition in dynamic scenes hinders robust mapping.

Purpose of the Study:

• To propose a novel dynamic SLAM algorithm leveraging deep learning for improved accuracy in dynamic environments.
• To enhance the front-end of SLAM systems with semantic information for better scene understanding.
• To develop a robust static map construction method by filtering dynamic object features.

Main Methods:

• Integration of YOLOv10n for semantic information extraction from image frames.
• Utilizing ORB-SLAM2 for feature point extraction and semantic information retrieval.
• Implementing a map construction thread to eliminate dynamic object features and build a static map.

Main Results:

• The proposed dynamic SLAM algorithm achieved over 96% accuracy improvement compared to ORB-SLAM2 in highly dynamic environments.
• Demonstrated superior accuracy and runtime performance against existing dynamic SLAM algorithms.
• Successfully constructed static maps by effectively identifying and removing dynamic object features.

Conclusions:

• The deep learning-based dynamic SLAM algorithm significantly enhances localization and mapping accuracy in dynamic environments.
• The integration of YOLOv10n and ORB-SLAM2 provides a robust solution for semantic SLAM.
• This approach offers a promising direction for reliable autonomous navigation systems in complex, changing conditions.