DOT-SLAM: A Stereo Visual Simultaneous Localization and Mapping (SLAM) System with Dynamic Object Tracking Based on...

Area of Science:

• Robotics
• Computer Vision
• Autonomous Systems

Background:

• Visual SLAM systems often assume static environments, leading to reduced accuracy with dynamic objects.
• Dynamic objects, especially vehicles, significantly degrade localization performance in autonomous driving.

Purpose of the Study:

• To develop a novel stereo visual SLAM system (DOT-SLAM) that integrates dynamic object tracking.
• To enhance ego-vehicle localization accuracy and generate static environment maps in the presence of dynamic objects.

Main Methods:

• Integrated dynamic object pose estimation using graph optimization.
• Employed a coarse-to-fine depth estimation method leveraging camera-road plane geometry.
• Utilized road plane and non-holonomic constraints for accurate dynamic object initialization.
• Jointly optimized ego-vehicle pose, dynamic object poses, foreground/background points, and road plane via graph optimization.

Main Results:

• DOT-SLAM effectively utilizes features from both dynamic objects and the background.
• Achieved more accurate vehicle trajectory estimation compared to traditional methods.
• Generated a static environment map, demonstrating improved localization performance.
• Validated effectiveness on the KITTI-360 dataset and real-world tests.

Conclusions:

• DOT-SLAM significantly improves localization accuracy in dynamic environments for autonomous vehicles.
• The system successfully integrates dynamic object information for robust SLAM.
• Offers a promising solution for reliable navigation in complex, real-world scenarios.