Robust place recognition under illumination changes using pseudo-LiDAR from omnidirectional images.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Visual Place Recognition (VPR) systems struggle with appearance changes due to illumination and sensor variations.
• Existing methods lack robustness in dynamic and heterogeneous visual environments.

Purpose of the Study:

• To develop a novel framework for robust Visual Place Recognition (VPR) resilient to illumination changes and sensor heterogeneity.
• To leverage depth estimation for enhanced scene representation in VPR systems.
• To provide a cost-effective alternative to 3D sensors for robot localization.

Main Methods:

• Transforming omnidirectional images into depth maps using Distill Any Depth (based on Depth Anything V2).
• Converting depth maps into pseudo-LiDAR point clouds.
• Utilizing the MinkUNeXt architecture for global-appearance descriptor generation.
• Implementing a novel data augmentation technique with distilled depth models.

Main Results:

• The proposed VPR system demonstrates robust performance across diverse lighting conditions, even after training on limited cloudy data.
• The framework generalizes well to geometrically dissimilar inputs, validated on different datasets and camera types.
• The approach shows competitive performance against state-of-the-art methods, especially in challenging illumination scenarios.

Conclusions:

• Depth estimation-based scene representation offers a robust solution for VPR challenges.
• Pseudo-LiDAR generation from standard cameras is a viable and cost-effective alternative to 3D sensors.
• The framework has significant potential for improving robot localization in complex environments.