Optimizing Appearance-Based Localization with Catadioptric Cameras: Small-Footprint Models for Real-Time Inference on...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Appearance-based localization is crucial for mobile robot navigation.
• Catadioptric cameras provide omnidirectional views but introduce image distortions.
• Recognizing places in visually similar indoor environments with limited features is challenging.

Purpose of the Study:

• To design an efficient neural network architecture for appearance-based localization using catadioptric images.
• To develop small-footprint models for real-time inference on edge devices for low-cost service robots.
• To evaluate the system's accuracy and efficiency compared to state-of-the-art methods.

Main Methods:

• Designing and comparing neural network architectures for visual place recognition.
• Utilizing transfer learning and fine-tuning on catadioptric images to generate global descriptors (embeddings).
• Testing on custom datasets and publicly available datasets (COLD Freiburg, Saarbrücken).

Main Results:

• The best results were achieved using embeddings from transfer learning and fine-tuning.
• The proposed system demonstrated favorable accuracy in place recognition.
• The system achieved competitive inference times, suitable for real-time applications.

Conclusions:

• The developed system offers a cost- and energy-efficient solution for appearance-based localization.
• The approach is effective for indoor service robots using catadioptric cameras.
• The optimized neural network architecture enables reliable localization even in challenging environments.