This paper presents a novel dynamic visual-inertial SLAM framework designed to enhance localization and mapping robustness in highly dynamic environments. Unlike traditional approaches that assume static scenes, our method integrates a scene flow-based dynamic object discrimination algorithm with Kullback-Leibler divergence thresholding to differentiate between dynamic and potentially dynamic objects. This strategy effectively preserves valuable feature points on temporarily stationary objects that are often mistakenly discarded by conventional methods. Furthermore, an efficient line feature filtering strategy is introduced, and combined with point features, it compensates for the scarcity of reliable point features in dynamic settings. Built upon the VINS-Mono platform and utilizing YOLOX for object detection, the framework is rigorously validated on the KITTI and ADVIO datasets. Experimental results demonstrate that our approach achieves an average ATE of only 1.23 m, significantly outperforming conventional methods. This strongly validates that our integration of Scene flow with KL divergence for dynamic judgment effectively enhances the robustness of SLAM in dynamic environments.