| Efficient and complete restoration of polluted sites is the significant guarantee for protecting the urban environment and citizens’ safety.As the improvement in the safety of the site construction personnel are needed,it is meaningful to use the human pose estimation technology to monitor their personal health status by real-timely detecting and analyzing their movements and postures during the work of hazard control and land restoration in the contaminated site.As the research related to computer vision technology and deep learning are developing,progressive findings of the human pose estimation method based on convolutional neural network are discussed.However,problems in practical application deployment are non-ignorable.On the one hand,how to make good use of image multiscale information to filter out complex background interference while retaining the key human body information in the picture.On the other hand,most of the current body pose estimation studies focus on the improvement of accuracy,but ignore the efficiency.This research studies in perspective of high-precision positioning and high-efficiency deployment,the solution of those two questions are discussed.The contributed findings of this research are concluded as follows:(1)To tackle the problems of complex background and non-rigid human pose in the pollution remediation site,this study puts forward the idea of the great sensing field hourglass attention network,designing the great sensing field residual module and the improved residual attention module to improve the traditional residual module and the original skip connection structure in the hourglass network.By expanding the area of effective receptive field,the great sensing field residual module helps the model to make good use of image multi-scale information,which could accurately locate body parts and joints to improve the accuracy and robustness of the human pose estimation.The improved residual attention module can effectively retain key human information in the image and filter out complex background interference by adding masks to human body areas.Finally,the research results that the average detection accuracy of PCKh@0.5 on the MPII data set is 91.8.The average accuracy of MSCOCO data set is improved by 4.3% over g-RMI model and 0.5% over RMPE model.(2)For the low efficiency of the existing human pose estimation models,an efficient human pose estimation network architecture based on hierarchical context learning was presented.Based on the preattentional processing mechanism in human visual system,a multi-stage hierarchical context network is constructed.In each stage of hierarchical context architecture,the trunk network adopts the large sense field hourglass attention network with different complexity,which concludes that the model can retain strong generalization ability while training and deploying efficiently.The experimental findings present that the average detection accuracy of PCKh@0.5 in the MPII data set is 91.7 in the hierarchical context hourglass network,and the computational overhead of the model is 260 P-flops lower than that of the stacked hourglass network. |
