DMS-SLAM: A General Visual SLAM System for Dynamic Scenes with Multiple Sensors.

Area of Science:

• Robotics and Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Existing SLAM systems excel in static environments but struggle with dynamic objects, leading to pose estimation errors.
• Dynamic elements in scenes pose significant challenges for mobile robots and autonomous driving applications.

Purpose of the Study:

• To propose a general visual SLAM system, DMS-SLAM, capable of operating effectively in dynamic scenes.
• To enhance pose tracking, loop closure detection, and relocalization accuracy in the presence of moving objects.

Main Methods:

• System initialization using Graph-based Motion Segmentation (GMS) and sliding window to create a static 3D map.
• Reprojection and constant speed/reference frame models for pose estimation and 3D map point updates.
• GMS feature matching for incorporating static 3D map points from selected keyframes.

Main Results:

• DMS-SLAM successfully implements pose tracking, loop closure, and relocalization in dynamic environments.
• The system supports various visual sensors including monocular, stereo, and RGB-D.
• Evaluations on TUM and KITTI datasets show superior accuracy and speed compared to state-of-the-art SLAM systems.

Conclusions:

• DMS-SLAM provides a robust solution for visual SLAM in dynamic scenes.
• The proposed method effectively mitigates the impact of dynamic objects on localization and mapping.
• DMS-SLAM offers improved performance for applications like mobile robotics and autonomous driving.