In order to improve the intelligence of the shop floor monitoring system and the accuracy of pedestrian detection in working scenes, a method based on computer vision technique is proposed. The task was simplified to regression prediction of centroids and scales by an anchor-free feature detection technique based on advanced semantic information. The feature extraction module obtained multi-scale image features applying 4-stage downsampling convolutional network and fused them. The head detection module was divided into two convolutions to process the feature maps in parallel to obtain centroid heat map and scale information and output the detection results. The results show that the MR-2 on the CityPersons dataset R subset reaches 11.61%, and the MR-2 is improved by 0.6% after adding the offset prediction branch. This proves the excellent performance of the personnel testing method.