Abstract
Recent advancements in 3D Gaussian Splatting (3DGS) have significantly improved the mapping quality and computational efficiency of visual Simultaneous Localization and Mapping (SLAM). We propose DOGL-SLAM, a novel framework that integrates 3DGS into its core pipeline, enabling accurate camera pose tracking, object-level interaction, and high-fidelity scene reconstruction in dynamic environments. Firstly, a joint graph optimization module incorporates both dense Gaussian and sparse landmark constraints, enabling precise alignment between camera tracking and mapping. Secondly, a consistent object-level semantic fusion module embeds category labels into 3D Gaussians, grouping them via semantic consistency and optimizing distributions through cross-view losses to support scene manipulation. Finally, a hierarchical dynamic filtering pipeline is introduced, consisting of segmentation, dynamic feature exclusion, gradient masking, and visibility-aware Gaussian pruning across three parallel threads. Our system is evaluated across multiple datasets, showing a significant improvement in high-quality view synthesis for dynamic scenes. Additionally, the generated object-level semantic maps facilitate advanced downstream tasks, highlighting the potential of a robust SLAM framework. The source code and supplementary video will be released.
Funding Information
  • National Natural Science Foundation of China (62303249)
  • China Postdoctoral Science Foundation-Tianjin Joint Support Program (2023T013TJ)
  • China Postdoctoral Science Foundation (2024M751526)
  • Beijing-Tianjin-Hebei Fundamental Research Cooperation Project (24JCZXJC00390)