Stereo Visual Odometry Pose Correction through Unsupervised Deep Learning.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual simultaneous localization and mapping (VSLAM) is crucial for robot positioning and navigation.
• Classical visual odometry (VO) systems struggle in challenging environments due to violated modeling assumptions.
• Deep learning offers robustness but often requires ground truth data for training.

Purpose of the Study:

• To develop an unsupervised pose correction network for classical stereo VO systems.
• To improve the accuracy and robustness of VO in challenging environments.
• To enable accurate VSLAM without relying on ground truth datasets.

Main Methods:

• Combined multiview geometry constraints with deep learning.
• Developed an unsupervised pose correction network for stereo VO.
• Network regresses pose corrections and generates depth maps and explainability masks.
• Trained the network without ground truth pose data.

Main Results:

• Significantly improved positioning accuracy of classical stereo VO.
• Achieved average absolute trajectory error of 13.77 cm over 100-800m on the KITTI dataset.
• Demonstrated competitive performance, nearing state-of-the-art results.
• Successfully generated depth maps and explainability masks simultaneously.

Conclusions:

• The unsupervised pose correction network effectively enhances stereo VO accuracy.
• The proposed method offers a robust solution for VSLAM in challenging conditions.
• This approach advances the field of robot localization and navigation.