Robust Visual Odometry Leveraging Mixture of Manhattan Frames in Indoor Environments.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Existing Visual Odometry (VO) and Simultaneous Localization and Mapping (SLAM) systems often rely on the Manhattan World (MW) assumption, limiting their applicability in diverse indoor environments.
• Previous methods struggle with drift and computational complexity, hindering real-time performance and accuracy in non-ideal indoor scenes.

Purpose of the Study:

• To develop a robust RGB-Depth Visual Odometry (RGB-D VO) system capable of accurate indoor localization.
• To overcome the limitations of the Manhattan World assumption by handling both MW and non-MW indoor scenes.
• To improve localization accuracy and reduce computational complexity compared to state-of-the-art methods.

Main Methods:

• A novel approach for detecting Manhattan Frames (MFs) using dominant directions from parallel lines, offering lower computational complexity than plane-based methods.
• Separate estimation of rotational and translational motion for MW scenes, utilizing MF observations, line direction vectors, and surface normals for drift-free rotation.
• Integration of tracked dominant directions with point and line features for full 6-DoF pose estimation in non-MW scenes, enhanced by multi-view rotation constraints and bundle adjustment.

Main Results:

• Achieved an average localization accuracy of 1.5 cm on synthesized datasets, comparable to state-of-the-art methods.
• Demonstrated superior performance on real-world datasets with an average localization accuracy of 1.7 cm, outperforming existing methods by 43%.
• Reduced time consumption by 36% compared to previous techniques.

Conclusions:

• The proposed RGB-D VO system effectively improves indoor localization accuracy by leveraging geometric features and adapting to different indoor scene types.
• The method offers a significant reduction in computational cost while maintaining or exceeding the performance of current leading approaches.
• This work provides a more versatile and efficient solution for camera pose estimation in complex indoor environments.