Deep features based real-time SLAM

Abstract

This project implements a near real-time stereo SLAM system designed to operate effectively in extreme conditions using Deep Learning methods. It employs a Parallel Tracking-and-Mapping approach, making use of stereo constraints to ensure robust initialization and accurate scale recovery while maintaining real-time performance. To handle various real-world challenges including dynamic illumination variations, the system integrates Convolutional Neural Networks (CNN) and Graph Neural Networks (GNN) for reliable corner point detection and matching. By optimizing the developed pipeline and integrating with CNN and GNN components, the system achieves near real-time performance.

Evaluations across diverse datasets with varying illumination conditions demonstrated the developed system's superiority over traditional feature-based methods in terms of accuracy and robustness. Notably, the system's implementation in Python prioritizes extensibility, making it both easy to read and understand, at the same time encourages customization and further development in terms of research, hence it potentially fosters progress in SLAM systems for various applications. Furthermore, the project explores the system's adaptability in underwater contexts, showcasing its workability even in extreme real-world scenarios.