Robust Depth Estimation Under Sensor Degradations: A Multi-Sensor Fusion Perspective.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Depth estimation is crucial for various applications.
• Current Multi-Sensor Fusion (MSF) methods struggle with sensor degradations, especially Out-of-Distribution (OOD) types.
• Data-driven learning is insufficient for OOD sensor degradations.

Purpose of the Study:

• To propose a novel approach, Combinable and Separable Multi-Sensor Fusion (CSMSF), for robust depth estimation.
• To enhance resilience against multiple sensor degradations.
• To enable autonomous selection of valid sensors for depth estimation.

Main Methods:

• CSMSF operates on four core principles: performance increases with valid sensors, independent estimation from a single sensor, balancing accuracy and complexity, and autonomous failure diagnosis.
• The method identifies and rejects degraded sensors.
• Valid sensors are autonomously selected for scene depth estimation.

Main Results:

• CSMSF demonstrates superior robustness against various sensor degradations.
• The approach maintains accuracy and resilience in diverse environmental conditions.
• Experimental results validate the efficacy of CSMSF.

Conclusions:

• CSMSF significantly improves depth estimation robustness.
• The proposed method effectively addresses challenges posed by sensor degradations.
• CSMSF offers a promising solution for reliable depth estimation in real-world scenarios.