Research Of Human Key Points Detection Based On Deep Convolutional Network

Human body keypoint detection is a technology based on image or video to locate joints.It has a wide range of applications in computer vision,such as behavior detection,pedestrian re-recognition,pedestrian behavior prediction and human-computer interaction.With the introduction of deep learning into the field of human keypoint detection,the accuracy and robustness of human keypoint detection have been greatly improved in recent years.However,it is more and more difficult to apply the human keypoint detection network to the practical application because the universal network adopts the structure design of a large number of stacked network layers,which results in a large increase in the number of parameters and computation of the whole network.For the above problems,this thesis studied the method of convolution modules and residual module structure,combined with the attention mechanism redesigned the residual module,key points in the human body detection accuracy were similar,effectively reduce the network number and amount of calculation,greatly reduces the network operation need calculate cost of storage space and network model.The main work of this thesis is summarized as follows:(1)The improved method of lightweight model,the structure of residual module and bottleneck module are introduced.In order to solve the problems of excessive number of parameters and high computational complexity caused by the high resolution representation of deep convolutional network,proposed two lightweight basic network modules,namely Gattneck module and Gattblock module.Based on HRNet(High-Resolution Network),we constructed a lightweight human key point detection network GATTNET(Ghost Attention Network).By introducing linear transformation to generate redundant feature graph and introducing channel attention mechanism which can reallocate channel weight,the lightweight improvement of HRNet is realized.The validation was performed on the MS COCO(Microsoft Common Objects in Context)2017 dataset,and the Experimental results show that the proposed GattNet network can effectively reduce the number of parameters and the complexity of operation while preserving the accuracy.(2)Inception network module design method and the classic channel attention module SE module are introduced.Inception attention module is designed by combining the two different design ideas.On the basis of(1),Inception attention module is added to the fusion part of GattNet network to enhance the extraction effect of features of different scales at the same resolution in the fusion stage,so as to achieve the realization that when the number of parameters is very small,the fusion effect can be improved.Finally validated on MS COCO 2017 data sets,the experimental results show that the proposed network in keeping with(1)a reference number on the basis of the accuracy is improved,significantly improve the effect of the integration of fusion module,on the basis of the realization of lightweight finally realize the improvement of the accuracy,is almost the same with the advanced network accuracy.