A Robust Semi-Direct 3D SLAM for Mobile Robot Based on Dense Optical Flow in Dynamic Scenes

Area of Science:

• Robotics
• Computer Vision

Background:

• Dynamic objects in scenes cause significant errors in mobile robot pose estimation.
• Inconsistent mapping due to dynamic objects hinders robot navigation and environmental understanding.

Purpose of the Study:

• To develop a robust semi-direct 3D simultaneous localization and mapping (SLAM) algorithm for mobile robots operating in dynamic environments.
• To enhance the accuracy and consistency of 3D maps built by mobile robots.

Main Methods:

• Utilizes a sparse direct method with homography matrix compensation for initial pose estimation.
• Employs dense optical flow to identify and eliminate dynamic regions, reducing error accumulation.
• Optimizes robot pose by minimizing reprojection error and incorporates a keyframe selection strategy.
• Performs global bundle adjustment (BA) for constructing a globally consistent 3D dense octree map.

Main Results:

• The algorithm successfully compensates for image deformation caused by robot rotation.
• Dynamic regions are effectively segmented and their influence on mapping is mitigated.
• Robot pose optimization is achieved by minimizing reprojection errors.
• A globally consistent 3D dense octree map is constructed, demonstrating superior performance.

Conclusions:

• The proposed semi-direct 3D SLAM algorithm offers a robust solution for mobile robots in dynamic scenes.
• The method significantly improves the accuracy and consistency of 3D mapping by addressing errors from dynamic objects.
• The algorithm's effectiveness is validated through simulations and experiments, showing superior performance.