MCVL: Multi-directional Camera-based Visual Localization Resistant to Seasonal and Illumination Changes

The global navigation satellite system-based technology has inherent limitations due to its reliance on radio signals. In contrast, visual localization operates independently of radio communication, presenting a viable solution to overcome these limitations. However, it is susceptible to seasonal and illumination variations, highlighting the need for research to address these challenges. Therefore, this paper proposes the multi-directional camera-based visual localization, which is robust against seasonal and illumination changes. The proposed method combines image from multiple directions and extracts global deep learning features. Subsequently, local deep learning features are extracted to preserve the characteristics of each combined image, allowing for the identification of geographically similar images. This approach utilize multi-directional cameras, enabling resilient performance under various constraints. Moreover, it demonstrates an improvement of 7.56% in recall rate at 1-meter threshold compared to existing methods.