The current research on simultaneous localization and mapping (SLAM) in academia mostly assumes static scenes, but dynamic objects are inevitable in real-life scenarios. Integrating deep learning into visual SLAM systems can collaboratively eliminate dynamic objects from the scene, effectively enhancing the robustness of visual SLAM in dynamic environments. This paper first introduces a classification of deep learning-based visual SLAM in dynamic environments and then provides a detailed overview of visual SLAM systems based on object detection, semantic segmentation, and instance segmentation. A comparative analysis of these approaches is also presented. Finally, considering the recent trends in the development of visual SLAM, the paper analyzes the main challenges of deep learning-based visual SLAM in dynamic environments and summarizes potential future directions.