In recent years, behavior recognition technology based on human pose information has been increasingly applied to worker safety monitoring. However, in complex industrial settings, challenges such as occlusion and limited compu-tational resources have hindered existing computer vision-based human pose estimation methods from simultaneously achieving high accuracy and low computational complexity. Thus, this paper proposes an approach for rapid human pose estimation in complex industrial scenarios by integrating a quantization autoencoder with a lightweight ResNeSt network. Furthermore, a cyclic weight transfer training method is proposed, which enhances estimation accuracy by transferring weight parameters across backbone networks of varying sizes. The experimental results demonstrate that the proposed method achieves accurate human pose estimation in complex industrial environments, reducing the computational cost of the original model by a factor of four, thereby providing an efficient and resource-effective solution for real-time pose estimation in industrial applications.