Research On Multi-person Pose Estimation Algorithm Based On Drilling Operation Site

As the basis of understanding the human body movements in the image,human pose estimation can locate the information of the keypoints of the human body from the image,so as to better analyze the human behavior.Multi-person pose estimation based on drilling operation site needs to distinguish and match multi-person poses in the drilling operation site,which makes pose estimation more challenging.On the one hand,the number of people in the drilling image is uncertain,so it is necessary to use the human object detector to traverse the image to know the number of people in the image,which puts forward higher requirements for the speed and accuracy of the human object detector.On the other hand,the camera with variable angle,the occlusion of other objects or people to the human joint points,and the complex background of the drilling site increases the complexity of the key point positioning.To solve the above problems,this paper proposes an improved Channel Shuffle and Attention Residual Stacked Hourglass Networks(CA-SHN)based on Stacked Hourglass Networks,which can effectively improve the accuracy of multi-person pose estimation.The main work of this paper is as follows:(1)An improved human object detection model(YOLOv3-person)based on YOLOv3 is proposed.According to the specific dataset of drilling operation site,the K-means clustering algorithm is used to design the appropriate anchor frame,Soft-NMS is used to improve the recall rate in the case of multi-person overlapping,the small pixel human object is ignored and the Spatial Transformer Network is used to adjust the position of the human object to improve the detection accuracy.(2)Based on the top-down framework and the improved human object detection model,this paper proposes an improved multi-person pose estimation algorithm CA-SHN based on Stacked Hourglass Network.Aiming at the problem that the recognition effect of small-scale joint points is not good in the case of overlapping of multi-person,multi-scale fusion is carried out in the channel dimension,the Conv＿1-Conv＿4 from downsample of the Hourglass Network are shuffled on the channel,so that the low-level features and high-level features complement each other.At the same time,the Residual Unit and Attention Mechanism are combined to form a Spatial & Channel Attention Residual Module,which can reduce the interference of complex background,improve the network attention to small-scale keypoints,and then improve the recognition effect of human pose estimation.Experiments on common datasets and self-built drilling field data show that CA-SHN algorithm improves the detection results of multi-person pose estimation.(3)Design the algorithm of worker behavior recognition based on pose estimation information.Aiming at the specific illegal behaviors of workers in the drilling operation site,the VGG-16 classification network combines with position information of the keypoints of human body,it can automatically identify the illegal actions of workers,such as crossing railings,dumping waste liquid,knocking operation,smoking and so on,which provides a guarantee for the safety production of the drilling operation site.