Kalman Filter based Image Re-selection: Handling Collinear and High Co-visibility Situations in Image Retrieval based Visual Localization

Autonomous driving systems must maintain consistent position estimation across various environments. However, conventional GPS-based position estimation technologies often yield inaccurate results in environments with weak signal reception, necessitating alternative solutions. Image retrieval based visual localization, unaffected by radio wave conditions, presents a robust alternative for accurate position estimation. Image retrieval, widely utilized in visual localization research, typically calculates precise positions by incorporating relative poses. However, accuracy may diminish in collinear and high co-visibility scenarios, requiring supplementary approaches. This paper proposes a method to overcome collinear and high co-visibility situations encountered in image retrieval. The proposed method applies retrieval results to a Kalman filter based system. Compared to conventional methods, the proposed approach demonstrated an average improvement of 11.12 meters in GPS error.