<p>3D human pose estimation (3D HPE) is a fundamental problem in applications such as action understanding, human–computer interaction, and virtual reality. The 2D-to-3D lifting paradigm is widely adopted because it provides a streamlined pipeline and controllable inference overhead. However, monocular estimation still suffers from inherent depth ambiguity due to single-view projection and from self-occlusion caused by human body articulation. As a result, complex models are often required to capture long-range, global dependencies. To balance accuracy and efficiency, this paper proposes a lightweight Spectral-Hypergraph Fusion Network (SHFPose). We introduce a Spectral Frequency Enhancement Module (SFEM) that enforces global geometric consistency via frequency-domain filtering. Concurrently, we design a Pose-Aware Soft Hypergraph (PASH) module that uses dynamic multi-hop routing to capture higher-order topological relations. This design complements the limited receptive field of local neighborhood propagation when modeling long-range dependencies. Features from these two branches are adaptively fused through a symmetric gating mechanism, which improves the structural stability of pose regression within a lightweight computational budget. Experiments on the Human3.6M and MPI-INF-3DHP benchmarks show that SHFPose achieves competitive performance across standard metrics while maintaining a low parameter count and computational cost. Runtime evaluations indicate its potential for real-time inference under resource-constrained settings. Overall, this method provides an efficient approach for modeling structural priors in single-image 3D human pose estimation, and shows potential value for real-time and edge-oriented applications.</p>