Abstract: Dynamic scene reconstruction, a key research direction in computer graphics and vision, aims to accurately capture and recover the three-dimensional geometric structure and time-varying appearance of dynamic scenes from monocular or multi-view two-dimensional image sequences or videos. 3D Gaussian splatting (3DGS) is a scene reconstruction and rendering technique based on three-dimensional Gaussian volumes. The parameterized nature of Gaussian volumes allows them to both explicitly represent scene elements and implicitly optimize local details, effectively simulating light propagation and scattering in the scene, ultimately enabling real-time rendering. This article reviews dynamic scene reconstruction methods based on 3DGS and categorizes them into three main categories. First, the basic principles and advantages of 3DGS are introduced, and related applications are briefly described. Second, three categories of methods are summarized: deformation fields, spatiotemporal four-dimensional Gaussian volumes, and inter-frame dynamic transfer. Various works are described, focusing on scene representation, motion modeling, and training strategies. The advantages and disadvantages of each category are compared and analyzed. Finally, the current status of development in this field is reviewed, and potential future research directions are identified. This includes the design of spatiotemporal data structures, dynamic reconstruction under sparse viewports, and real-time reconstruction of large-scale dynamic scenes.