Amid the trends of industrial automation and intelligent development, human action recognition technology provides full-process support encompassing “real-time monitoring-early warning-intervention”. It transforms factory work safety management from “post-accident remedy” approach to “pre-accident prevention”, which not only safeguards workers’ lives but also helps enterprises reduce accident losses and improve production stability. To address the decline in action recognition accuracy caused by object occlusion of the human body in complex real-world factory scenarios, an occluded human pose estimation algorithm based on feature self-calibration and feature refinement was proposed. Firstly, a feature self-calibration unit (FSCU) was utilized to adaptively model and learn the dependency relationships between spatial and channel dimensions. While generating a global receptive field, it reduced the impact of irrelevant regions through channel information, thereby obtaining more discriminative feature representations of joint points. Subsequently, a multi-granularity feature refinement unit (MGFRU) was designed, which maintained high resolution in both spatial and channel dimensions and focuses on perceiving key information in different regions. This unit filters and refines multi-granularity features extracted by the backbone network, highlighting key features related to joint points. The model proposed was analyzed and compared with mainstream models, the proposed model achieves an overall mean accuracy of 90.8% on the MPII public dataset, outperforming all other listed state-of-the-art models. On the self-constructed factory scenario dataset that simulates different occlusion levels, this model exhibits stronger robustness, thus improving production safety and preventive reliability in practical engineering applications.