Enhancing Visual Localization Accuracy in Dynamic Environments via Masking Techniques

In recent years, with the advancement of deep learning, Visual Localization has played a crucial role in various application domains. Visual Localization refers to the ability of a system to accurately determine its own position, enabling safe and efficient navigation. Typically, positioning technologies such as GPS are used for location determination. however, the accuracy of these systems can degrade in environments where GPS signals are weak or unavailable, such as indoors. In such environments, methods that combine multiple sensors, such as accelerometers, gyroscopes, and cameras, prove to be effective. This approach allows for accurate position estimation even in conditions where GPS signals are unstable. However, environmental factors such as seasonal changes and vegetation can lead to a decline in accuracy, which remains a challenge to address. To tackle this problem, this study proposes a mask-based approach utilizing color detection and the Segment Anything Model (SAM), combined with Hierarchical Localization (HLoc) to achieve high-precision Visual Localization. We conduct comparative analyses of localization performance with and without masking under two different imaging conditions. Additionally, to ensure practical applicability, we develop a scale adjustment and pose compensation algorithm and quantitatively evaluate localization errors using a 3D model constructed with COLMAP. Experimental results indicate that under certain environmental conditions (such as cloudy weather), mask processing can improve the accuracy of position and pose estimation.