Visual Localization in Complex Environments: Merging Traditional Geometry with Learning-Based Techniques

This paper presents a hybrid visual localization approach that combines classical geometry-based methods with state-of-the-art deep learning techniques to improve localization accuracy in complex and dynamic environments. By leveraging feature matching techniques, multi-view integration and integrating semantic segmentation, our method achieves robust feature matches, the effects of the different is evaluated. The evaluation is done on real-world vehicle data with LiDAR, GNSS/IMU ground truth motion and cameras, the system demonstrates enhanced precision in camera pose estimation. These capabilities are critical for maintaining accurate, real-time spatial understanding in applications like autonomous driving and augmented reality. The proposed method also supports efficient updating of scene representations, laying the groundwork for the dynamic maintenance of digital twins essential in XR environments and the influence of semantic segmentation and camera geometries is evaluated on a challenging real world data set.