A Semantic SLAM System for Catadioptric Panoramic Cameras in Dynamic Environments

Area of Science:

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

Background:

• Traditional visual SLAM struggles in dynamic environments due to interference from moving objects.
• Catadioptric panoramic cameras offer a wider field of view but present unique challenges for SLAM.

Purpose of the Study:

• To develop a semantic SLAM system for catadioptric panoramic cameras that robustly handles dynamic environments.
• To improve the accuracy and reliability of pose estimation in the presence of dynamic objects.

Main Methods:

• Utilized a real-time instance segmentation network to detect potential dynamic targets in panoramic images.
• Applied sphere's epipolar constraints to verify and filter out actual dynamic objects.
• Masked dynamic objects during feature point extraction to focus on static elements for pose estimation.

Main Results:

• The proposed semantic SLAM system demonstrated significantly improved accuracy compared to traditional algorithms in highly dynamic environments.
• Achieved up to 96.3% better performance in terms of Root Mean Square Error (RMSE) of absolute trajectory error.
• The system proved to be more accurate and robust in complex dynamic environments.

Conclusions:

• The semantic SLAM system effectively overcomes the limitations of traditional SLAM in dynamic scenarios.
• Masking dynamic objects based on semantic information enhances the robustness and accuracy of pose estimation for panoramic cameras.
• This approach offers a significant advancement for autonomous systems operating in complex, real-world environments.