Abstract: Protective wearing detection and action recognition are an important part in the intelligent safety supervision system. Aiming at the problem of low accuracy and low real-time performance of traditional parallel recognition methods, We propose a parallel recognition scheme based on pose estimation to realize protective wearing detection and action recognition in parallel quickly, which uses a data interaction method based on distance and Hungarian algorithm to combine the pose with the guardian. Useing deep separable convolution to reduce parameters, the parallel recognition scheme improve the Openpose model, which makes the Openpose model lighter and improves the real-time performance of pose estimation. A quadrilateral detection method combined with pose information is proposed to solve the misjudgment problem of not wearing a guardian (holding helmet). In real-time detection experimental, the misjudgment rate of the holding helmet drops by 29.3%, and the overall accuracy rate of the system reaches 93.5%. The speed of improved Openpose model increases 12 frames per second faster than the original model, which is 2.2 times that of the original model. Experimental verification shows that the parallel recognition scheme has high accuracy and strong real-time performance, which meets the needs of actual protective wearing detection and action recognition.