Investigating the Influence of Scale Cues and Pose Integration on AI-Based Monocular Depth Estimation for Resource-Constrained Mobile Robots

Depth estimation is crucial for robotic navigation, with monocular depth estimators providing cost-effective and accessible solutions. However, their accuracy can degrade in atypical camera poses. This thesis investigates a novel approach to addressing pose biases in AI-based monocular depth estimation by incorporating camera poses obtained using scale-aware feature extraction as an additional input parameter. The methodology encodes the front-facing camera pose of a rover as an additional input channel to a U-Net-based monocular depth estimator. The pose is extracted from images captured by a rear-facing camera, typically used for regolith tracking, using SIFT and RANSAC algorithms. Scaling is performed using known dimensions of objects in the rear view, followed by refinement with a particle filter. Different pose encoding techniques are analyzed, highlighting their potential to improve depth estimation accuracy while identifying key areas for further optimization.