Research On Multi-human Skeleton Extraction Algorithm Based On Convolutional Neural Network

Skeleton extraction of the human body is a very important research direction in the field of computer vision research.With the increasing level of computer hardware and the exponential increase of computing power,there has been a great progress in image processing,so the demand for processing and analyzing human motion on computers are increasing.The human skeleton extraction technology is the basis for analyzing behavioral actions.This technology has very important practical application value.For example,in the field of monitoring,it is desirable to pre-determine whether the person's actions are dangerous.In sports and dance,it is also hoped to be carried out by computer.The human body action scores as a reference standard for the game.In recent years,with the development of big data and artificial intelligence,many areas of the problem have been transferred to deep learning convolutional neural networks.The extraction of the human skeleton has also been changed from the traditional method to the convolutional neural networks.In practical applications,the scene of multiple people is larger than the scene of single person,and the extraction of multi-human skeleton has more practical significance.In this paper,based on convolutional neural network,the scheme of extracting multi-human human skeleton is expounded in the two directions of static picture and video.The specific research contents are as follows:In the multi-human skeleton extraction of static images,this paper discusses the difference between the top-down and bottom-up methods,and uses the bottom-up method to identify the images,uses the heatmap to represent the feature extraction of bone nodes.Then the improved Part Affinity Fields method is used to find the relationship between bone nodes to optimize joint-to-human matching,and finally based on improved residual network as the basic model,using data on MPII dataset Training,get recognition and recognition rate.In the video-based multi-human skeleton extraction,this paper uses two identification schemes.In the first scheme,based on the static image recognition scheme,the RNN-based method is used to learn the frames between frames.The association of the nodes and the influence of the RNN improved model GRU on the recognition,and finally the recognition results and accuracy of the video are obtained.In the second scheme,the method of the Faster R-CNN is used to detect the human body frame in a single frame,and in each human body frame.The bone nodes of each human body are detected,and finally the IOU tracking method is used to find the relationship between the frame and the frame in the video to optimize the recognition effect.