Loop Closure Detection (LCD) is an essential component of visual simultaneous localization and mapping (SLAM) systems. It enables the recognition of previously visited scenes to eliminate pose and map estimate drifts arising from long-term exploration. However, current appearance-based LCD methods face significant challenges, including high computational costs, viewpoint variance, and dynamic objects in scenes. This paper introduces an online based on Superpixel Grids (SGs) LCD approach, SGIDN-LCD, to find similarities between scenes via hand-crafted features extracted from SGs. Unlike traditional Bag-of-Words (BoW) models requiring pre-training, we propose an adaptive mechanism to group similar images called $\textbf{\textit{dynamic}}$ $\textbf{\textit{node}}$, which incremental adjusts the database in an online manner, allowing for efficient retrieval of previously viewed images. Experimental results demonstrate the SGIDN-LCD significantly improving LCD precision-recall and efficiency. Moreover, our proposed overall LCD method outperforms state-of-the-art approaches on multiple typical datasets.