DyMoSLAM: Improving Dynamic Monocular SLAM With Neural Scene Representation

To address the significant positioning errors and inaccurate dense map construction of the traditional simultaneous localization and mapping (SLAM) framework when facing dynamic scenes, this article proposes a dynamic visual SLAM system, DyMoSLAM, based on neural implicit representation. The novelty of this system lies in the integration of a dynamic region detection mechanism, a mapping strategy based on implicit neural representation, and a detailed reconstruction of dynamic radiation field. By integrating the improved optical flow masks and instance segmentation techniques, DyMoSLAM effectively distinguishes between dynamic and static elements in the scene, thereby improving the accuracy of camera pose estimation. Additionally, we adopt a mixed encoding strategy combining triplane and one-blob representations to optimize memory usage and computational efficiency. Specialized dynamic and static loss functions are also introduced to enhance both mapping accuracy and detail restoration. Extensive experiments demonstrate that DyMoSLAM achieves an average tracking error of 3.42 cm on the dynamic sequences of the TUM RGB-D dataset, significantly outperforming existing NeRF-based SLAM systems. Moreover, it attains the best mesh completeness (1.92 cm) and highest completeness ratio (93.49%) on the replica dataset, indicating superior mapping quality in dynamic scenes. The proposed method shows strong potential for high-fidelity reconstruction and real-time localization in challenging dynamic environments.