Unsupervised Estimation of Monocular Depth and VO in Dynamic Environments via Hybrid Masks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Deep learning excels in 3-D sensing but struggles with dynamic environments.
• Existing monocular visual odometry (VO) methods are susceptible to failures caused by moving objects.

Purpose of the Study:

• To mitigate the negative impact of dynamic environments on joint depth and VO estimation.
• To improve the robustness and accuracy of 3-D perception systems.

Main Methods:

• Proposed hybrid masks (cover and filter masks) to address dynamic elements in VO and depth estimation.
• Introduced a depth-pose consistency loss to resolve scale inconsistencies in monocular sequences.
• Developed a joint estimation framework for depth and VO.

Main Results:

• Achieved state-of-the-art depth prediction and globally consistent VO estimation on the KITTI benchmark.
• Demonstrated the method's transferability on the Make3D dataset.
• Hybrid masks effectively reduced the adverse effects of dynamic environments.

Conclusions:

• The proposed hybrid mask approach significantly enhances deep learning-based 3-D sensing in dynamic scenarios.
• The method provides accurate and robust depth and VO estimation, crucial for autonomous systems.
• The depth-pose consistency loss further improves the reliability of monocular sequence analysis.