ScholarWorks@동국대학교

초록

Accurate and real-time 3D human pose estimation is essential for safety-critical applications such as intelligent surveillance, yet remains challenging in resource-constrained and dynamic environments due to its high computational demands. To address this, we propose a cooperative edge inference method for real-time 3D pose estimation in mobile edge computing networks. End devices equipped with lightweight models apply dual confidence thresholds to filter uncertain inputs, offloading only the selected images to an edge server for refined inference. We formulate a joint optimization problem to determine the optimal confidence thresholds and transmission times per device with the aim of minimizing the mean per-joint position error under end-to-end delay constraints. To evaluate the proposed method under controlled and repeatable multi-view conditions, we developed a virtual 3D simulation environment in Unity that mimics motion scenarios and provides accurate ground-truth pose data. Experimental results demonstrate a clear trade-off between accuracy and latency, and confirm that the proposed method significantly improves estimation accuracy while consistently meeting delay requirements. © 2025 IEEE.

키워드