Human Pose Estimation Based On Attention Detection

The purpose of human pose estimation is to locate the key joints of human body.The key points include the joints of head and limbs.Human pose estimation can help to label different parts of human body more precisely.Also,it can help to get current state of human body,analyze human behavior,and track human movement.Current human pose estimation systems include single person and multi-person pose estimation system.A single person pose estimation system combining with a human detector can complete multi-person pose estimation tasks.Thus,a better single person pose estimation algorithm can do benefits to both single person and multi-person pose estimation.A deep convolutional network is the key of single person pose estimation system.The multi-resolution fusing neural network is one of the mainstream networks.It focuses on how to fuse feature maps efficiently.Now there are works fusing two or more resolution feature maps and they got good results.But these methods ignore the difference between different resolutions.Also,current residual blocks take their inputs without difference,ignoring the key point of residual block is to enhance the deeper feature.Targeting at exploiting the extra information in current hidden layers,we proposed a novel multi-resolution fusing neural network based on attention detection.It has two features.First,we proposed a novel residual block.It contains attention detection,boundary clip,and initialization constraint.It makes the new residual blocks focus on the deep feature enhancement.Second,we proposed a novel multi-resolution fusing block.It contains map channel attention and resolution attention detection.It also has boundary clip and initialization constraint.It makes the fusing block always focus on the resolution that is waiting to be processed.We validated our results using single person pose dataset and multi-person pose dataset.The results of attention blocks show that the attention parameters match our boundary clip design and have learned better representation than original setting.The pose estimation results show that our new network outperform or equal to current mainstream networks.