| dc.description.abstract | We proposes a method that combines data augmentation techniques with semantic information to address the issue of increased positioning errors in visual localization caused by dynamic environments. Visual localization is crucial in applications such as autonomous vehicles, robotics, and augmented reality (AR) / virtual reality (VR). However, in dynamic environments, especially where there is frequent human movement, localization accuracy and stability often significantly decline.
To solve this problem, we adopted the Random Erasing technique from data augmentation. Random Erasing simulates object movement or occlusion by randomly masking parts of the image, allowing the model to learn more diverse features and improve its robustness in dynamic environments. However, we believe the model should learn more useful features. Therefore, we further integrated semantic segmentation techniques to extract human regions in the images and applied special processing to these areas.
This combined approach aims to enhance the model′s adaptability in dynamic environments, ensuring localization accuracy in practical applications.
We conducted experiments on our datasets with varying dynamic characteristics in indoor environment like factory. Experimental results show that this method reduces localization errors caused by human movement. In areas with human movement, our method reduces translation errors by at least 35.8 \% and improves system stability. Additionally, in static environments, our method maintains high accuracy, demonstrating its adaptability across various settings. | en_US |