Deep Learning-Based Keypoints Driven VIO for GNSS-Denied Flight

This project aims to investigate the feasibility of replacing the conventional hand-crafted feature extractor module in visual-inertial odometry with a deep learning-based feature extractor. Our research demonstrates the comparable performance of the proposed method to the traditional one. Specifically, we implemented a hybrid system based on the widely used state-of-the-art visual-inertial system, VINS-Mono, by integrating convolutional neural network(CNN)-based features. Our experimental results indicate that the proposed method achieves performance on par with the conventional approach while offering potential advantages in terms of robustness and adaptability.