ADM-SLAM: Accurate and Fast Dynamic Visual SLAM with Adaptive Feature Point Extraction, Deeplabv3pro, and Multi-View...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual Simultaneous Localization and Mapping (V-SLAM) is essential for autonomous systems.
• Dynamic environments pose significant challenges for V-SLAM accuracy.
• Current deep learning methods for dynamic object recognition are computationally intensive.

Purpose of the Study:

• To develop an efficient V-SLAM system for dynamic environments.
• To overcome the computational limitations of existing dynamic object recognition models.
• To improve the accuracy and real-time performance of V-SLAM in challenging scenarios.

Main Methods:

• Proposed ADM-SLAM system building upon ORB-SLAM2.
• Integrated adaptive feature point homogenization.
• Employed lightweight deep learning semantic segmentation (improved DeepLabv3).
• Utilized multi-view geometric segmentation for motion state detection.

Main Results:

• ADM-SLAM significantly outperforms ORB-SLAM2 in dynamic environments.
• Achieved up to a 97% reduction in Absolute Trajectory Error (ATE) in high-dynamic scenes.
• Demonstrated superior real-time performance and accuracy compared to DS-SLAM and DynaSLAM.

Conclusions:

• ADM-SLAM effectively eliminates dynamic interference points.
• The system shows excellent adaptability and robustness in highly dynamic environments.
• Offers a viable solution for real-time V-SLAM in challenging conditions.