Dynamic Three-Dimensional (3D) human reconstruction represents a pivotal research direction at the intersection of computer vision and computer graphics, harboring substantial application potential in virtual reality and digital entertainment. Traditional reconstruction methodologies, primarily relying on explicit representations such as polygonal meshes, voxel grids, and point clouds, suffer from the limited representation capabilities inherent in underlying parametric models, making it difficult to capture complex non-rigid geometry, fast joint movements, and fine-grained texture details. Conversely, implicit reconstruction techniques, exemplified by Neural Radiance Fields (NeRF), are severely bottlenecked by prohibitive volume rendering computational overhead, which hinders real-time interaction and direct physical editing. In this context, 3D Gaussian Splatting (3DGS) has emerged as a revolutionary and highly promising paradigm. It leverages discrete, anisotropic 3D Gaussian primitives-characterized by attributes such as position, covariance, color, and opacity-, alongside an exceptionally efficient differentiable rasterization mechanism to establish a new frontier for high-fidelity, real-time dynamic human reconstruction. To elucidate the evolution of this technology, a universal baseline reconstruction pipeline was systematically abstracted to dissect the fundamental workflow into three core components: canonical space representation, pose-driven deformation fields, and tile-based differentiable rasterization. On this basis, mainstream improvements addressing three major research bottlenecks were analyzed and compared: First, regarding detail reconstruction enhancement, methods for enhancing the reconstruction of human details and complex garments; their integration of expressive parametric templates, multi-shell volumetric mappings, and semantic graph constraints to dramatically improve the geometric fidelity and texture details of high-degree-of-freedom areas like faces, hands, and multi-layered loose clothing. Second, in terms of computational resource management, methods for accelerating training and inference and reducing memory consumption; including hash grid encoding, keyframe-based initialization, adaptive Gaussian pruning, and neural attribute quantization-demonstrate collective efforts to reduce training and inference times to seconds and compress storage footprints to the sub-megabyte scale for mobile deployment. Third, concerning system robustness enhancement, meticulous analyses clarify how strategies employing generative diffusion priors, feedforward regression networks, material-geometry-lighting decoupling, and physical simulator bindings resolve complex challenges associated with severe self-occlusions, sparse inputs, models for improving robustness to occlusion, cross-identity generalization, relighting, and physical applications. Furthermore, the loss-function system, including reconstruction losses and geometric, deformation, and attribute regularization constraints, was comprehensively reviewed, and major datasets based on monocular video, multi-view video, and 3D scanning, together with key rendering evaluation metrics. Finally, promising future research avenues focus on the co-optimization of quality, efficiency, and robustness; multi-instance complex scene reconstruction; and advanced multimodal physical-semantic interaction driven by large foundation models. Ultimately, future directions were discussed to provide a systematic reference for subsequent studies.
