Loop Closure Detection Based On Semantic Information In Dynamic Scenes

This paper proposes an approach to optimize visual loop closure detection (LCD) algorithms for dynamic scenes. Although loop closure detection has been widely used in visual SLAM systems, there are still some deficiencies when running in real scenes, such as the vulnerability in scenes with dynamic targets, which severely interferes with the similarity calculation between images and greatly degrades the performance of loop closure detection. In addition, existing loop closure detection algorithms rely on traditional image features, making them more sensitive to lighting, rotation, etc., resulting in high sensitivity. Our approach aims to enhance the correlation between images by utilizing semantic information, improving the original feature matching algorithm based on the Bag-of-Words (BoW), and constructing a semantic matching algorithm based on the results of semantic data association, thereby improving the accuracy of loop closure detection in dynamic scenes. Extensive experiments on public datasets have demonstrated the reliability and effective-ness of our method, achieving better global localization accuracy compared to other loop closure detection algorithms.