Research On Human Pose Estimation Based On Attention Mechanism

Human pose estimation is an important research direction of computer vision.In recent years,with the development of deep learning technology,related research results are widely used in the fields of behavior detection,video surveillance,virtual reality and so on.However,the development of human pose estimation technology is still restricted by the bottleneck problems such as lack of global feature relationship,scale difference of target and occlusion of key points in practical application.For solving the key problems of human pose estimation technology,this paper proposes the High-Resolution Network,HRNet as a benchmark model,introduces attention mechanism from channel,space and scale,and integrates it into the benchmark model.On this basis,data enhancement is used to improve the performance of the model,so as to achieve the accurate extraction of the target key information.The main research contents of this paper are as follows:Firstly,in order to solve the problem that HRNet lacks global context in the output phase,a High-Resolution Network based on dual attention mechanism(DHRNet)is proposed by introducing parallel channel attention and spatial attention modules,so as to realize the rebalancing of output features in channel domain and spatial domain,obtain richer context.And cross attention mechanism is used to reduce the complexity of spatial attention module and improve the timeliness of the proposed method.Secondly,to solve the problem of insufficient scale information extraction and not weighing features between different scales in High-Resolution Network(HRNet),a High-Resolution Network based on scale attention(SHRNet)was proposed.From the perspectives of the same scale and different scales,a Recursive High-Resolution Network(RHR)which enhanced the same scale feature information and an attention structure SAM which weighed the different scale feature information were constructed respectively,so as to achieve the model performance improvement at the same scale and different scales.Lastly,in view of the lack of global feature relationship,difference of target scale and blocking of key points in the human body pose estimation method,dual attention mechanism and scale attention mechanism were used to improve HRNet.A high-resolution network based on attention mechanism AHRNET was proposed,and the performance changes of the improved model were evaluated through COCO data set.On this basis,the Grid Mask and Cutmix methods were used to enhance the data by simulating the occlusion of key points and the complex background to increase the diversity of data samples and improve the generalization ability of the model,so as to achieve the accurate extraction of the key information of multi-channel and multi-scale targets.All models in this paper are based on MPII data set and COCO data set for experiments and performance evaluation.