Parallax Inference for Robust Temporal Monocular Depth Estimation in Unstructured Environments.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Accurate depth estimation is critical for autonomous vehicles, but dedicated sensors face limitations.
• Estimating depth from onboard RGB cameras in natural outdoor environments presents significant challenges.

Purpose of the Study:

• To develop a novel depth estimation method for autonomous vehicles using only RGB cameras.
• To address the complexities of depth estimation in natural outdoor landscapes.

Main Methods:

• Established a bijective relationship between depth and visual parallax from consecutive frames.
• Developed a motion-invariant, pixel-wise depth estimation technique.
• Utilized a pyramidal convolutional neural network with customized cost volumes for refined parallax map estimation, leveraging spatio-temporal constraints.

Main Results:

• The proposed method achieves motion-invariant, pixel-wise depth estimation.
• The network demonstrated superior performance compared to state-of-the-art methods on public outdoor scene datasets.
• The approach showed robust performance across varied scenes and generalization capabilities.

Conclusions:

• The novel depth estimation method is effective for autonomous vehicles operating in natural outdoor landscapes.
• The technique offers a viable alternative to dedicated depth sensors under resource constraints.
• The pyramidal convolutional neural network architecture enhances robustness and accuracy in complex environments.