DSOMF: A Dynamic Environment Simultaneous Localization and Mapping Technique Based on Machine Learning.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Classical semantic simultaneous localization and mapping (SLAM) algorithms struggle with localization accuracy and map completeness in dynamic environments.
• Existing methods often fail to accurately track and map objects that change position or appearance over time.

Purpose of the Study:

• To develop a novel dynamic scene SLAM technique that enhances localization accuracy and map construction in real-time.
• To overcome the limitations of traditional SLAM in environments with significant object motion and appearance changes.

Main Methods:

• The proposed method builds upon Direct Sparse Odometry (DSO), integrating instance segmentation and video completion algorithms.
• It leverages semantic and geometric data fusion for robust object tracking and incorporates semantic probability to improve image matching.
• A loop closure module utilizing video inpainting on the static background is introduced to enhance localization reliability.

Main Results:

• The dynamic scene SLAM technique was validated on the TUM and KITTI datasets and an unmanned platform.
• Experimental results demonstrate significant improvements in localization accuracy, exceeding 85% compared to the baseline DSO system in various dynamic scenes.
• The method effectively handles dynamic elements by associating identical objects across frames and utilizing a completed static background for loop closure.

Conclusions:

• The developed dynamic scene SLAM approach significantly enhances localization accuracy and map quality in challenging dynamic environments.
• The integration of instance segmentation, video completion, and semantic data fusion offers a robust solution for real-time SLAM.
• This research provides a foundation for more reliable autonomous navigation systems operating in complex, real-world scenarios.