Performance Evaluation of Image-Aided Navigation with Deep-Learning Features

Visual Odometry (VO) and Visual Simultaneous Localization and Mapping (V-SLAM) are advanced techniques widely used in positioning and navigation applications. When combined with other sensors, such as GNSS and IMU, they provide a seamless and more robust positioning solution. Recently, deep-learning (DL) based approaches for tie point extraction, based on convolutional neural networks (CNN), are gaining popularity for image-based navigation, despite hand-crafted approaches that rely on classical local features remain widely used. Although the performance of DL-based methods is similar to hand-crafted local features in common scenarios, they tend to perform significantly better in challenging conditions, such as at significant illumination and perspective changes. However, their impact on real-time positioning is only at the beginning to be investigated. The contribution of the paper is threefold: (1) to evaluate the accuracy improvement potential offered by incorporating DL-based local features in image-aided navigation; (2) to examine the efficacy of these features in real-world contexts where GNSS signal is not available; (3) to report the advancements with an open-source pipeline, named COLMAP-SLAM, that facilitates the integration of diverse learning-based features to support image-based navigation.