Abstract: To address the issues of low rendering efficiency, rigid resource allocation, and insufficient dynamic adaptability of traditional dense visual SLAM algorithms in dynamic scenes, this paper proposes a dense visual SLAM algorithm based on dynamic adaptive 3D Gaussian splatting. The proposed method leverages explicit 3D Gaussian scene representation and differentiable splatting rendering techniques to construct a joint geometric-appearance optimization framework. A gradient-guided density control mechanism is employed to dynamically adjust the topological structure of Gaussian primitives, while a multimodal joint optimization strategy is incorporated to suppress noise interference and enhance the compactness and dynamic adaptability of scene reconstruction. For camera tracking, a motion-prior-guided robust tracking framework is designed, incorporating manifold extrapolation initialization and decoupled optimization strategies to improve the accuracy and stability of pose estimation in dynamic environments. Meanwhile, a keyframe-driven incremental optimization mechanism is adopted to effectively balance memory consumption and scene representation granularity. Experimental results demonstrate that the proposed method significantly outperforms existing approaches in dynamic complex scenes, achieving both high-precision tracking performance and real-time rendering efficiency with controllable resource usage, thereby providing an efficient and robust solution for dynamic environment perception in autonomous robot navigation and immersive AR/VR applications.