Real-time and accurate 3D human pose estimation remains challenging in intelligent surveillance networks due to limited computational and communication resources. To address this issue, we propose a cooperative inference method based on mobile edge computing (MEC), where end devices apply dual confidence thresholds to selectively offload images for refined server-side inference. By jointly optimizing device-specific confidence thresholds and transmission times, the proposed framework achieves a balanced tradeoff between accuracy and latency. Simulation results demonstrate that our method effectively minimizes the mean per-joint position error (MPJPE) while satisfying end-to-end delay constraints. The proposed approach provides a promising framework for effective, low-latency 3D human pose estimation in multi-device MEC environments.