Research On Human Pose Estimation Technology Based On Deep Learning

As a hot topic in the field of computer vision,human pose estimation has been widely used in intelligent security,human-computer interaction,health monitoring,motion recognition and other fields.The accuracy of attitude estimation is very important in many fields,and the lightweight of the model is also very important for some devices such as mobile terminals.Improving the accuracy and speed of human pose estimation algorithm is always a research hotspot in this field.Based on the accuracy and speed of the algorithm,the following studies are made in this paper:(1)This paper first introduces two classical human pose estimation algorithms,namely the bottom-up Open Pose algorithm and the top-down Alpha Pose algorithm.Through testing on MS COCO data set and MPII data set with high recognition accuracy and speed,Alpha Pose algorithm is obtained as the main research algorithm of this paper.(2)Secondly,aiming at the target detection network in AlphaPose algorithm,this paper studies the current mainstream target detection algorithm YOLOv4 and Mobile Net series lightweight backbone extraction network.Based on the idea of Mobile Net series networks and deep separable convolution,three lightweight YOLOv4 networks are created.Finally,it is verified by experiments that the number of model parameters is greatly reduced and the frame rate is greatly improved under the condition that the accuracy loss of the three networks is small compared with that of YOLOv4.(3)Aiming at the human pose estimation network in AlphaPose algorithm,this paper deeply studies the multi-scale feature fusion high-resolution network and channel spatial attention mechanism,and proposes a new high-resolution network SN-HRNET.Through testing on MS COCO dataset and MPII dataset,it is verified that the improved high resolution network can effectively improve the detection accuracy with little parameter number loss.(4)Finally,an improved Alpha Pose algorithm is designed using the lightweight YOLOv4 and SN-HRNET proposed in this paper,and the effectiveness of the improved algorithm is verified through a series of comparative experiments.