Deep Visual Odometry With Adaptive Memory.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Existing deep learning visual odometry (VO) methods often suffer from significant error accumulation due to treating VO as a pure tracking problem.
• Preserving global information is critical for mitigating these accumulated errors in end-to-end VO systems, but it remains a challenge.

Purpose of the Study:

• To develop a novel deep visual odometry (VO) method that effectively incorporates global information to overcome error accumulation.
• To enhance the robustness and accuracy of VO systems, particularly in challenging scenarios.

Main Methods:

• A novel deep visual odometry (VO) method is proposed, incorporating an adaptive memory module for progressive information saving (local to global) to process long-term dependencies.
• A refining module further enhances previous results using global information from memory.
• A spatial-temporal attention mechanism selects features based on co-visibility, guiding the refinement process.

Main Results:

• The proposed VO method demonstrates superior performance compared to state-of-the-art methods on the KITTI and TUM-RGBD datasets.
• It achieves competitive results against classic approaches in standard scenes.
• The model exhibits outstanding performance in challenging conditions like texture-less regions and abrupt motions, where traditional algorithms often fail.

Conclusions:

• The developed deep visual odometry (VO) method effectively leverages global information through memory and pose refinement, significantly reducing error accumulation.
• The approach offers a robust solution for accurate camera pose estimation, outperforming existing methods in both regular and challenging environments.