Real-Time Multi-Persons Pose Estimation In Complex Scenes

Human pose estimation is a process of restoring the location and connection of human joints from image or video,which is basic issue in computer vision and computer graphics.It plays an important role in many application fields,such as security monitoring,virtual reality,Action film and so on.At present,the single person pose estimation in simple scene is basically solved,but the human pose estimation in complex scene,especially the multi-person real-time pose estimation,is still a challenge.Human interaction occlusion,self-occlusion,dress diversification,illumination change and scene object interference exist in the complex scenes of real-time multi-human pose estimation,which leads to the multi-person attitude estimation problem has not been well solved.Problems in multi-person attitude estimation.This paper focuses on the key problems such as multi-scale feature extraction of human body,the error of human proposed bounding frame and the low efficiency of large-scale neural network.We proposes a new human pose estimation model.Firstly,in this paper,based on the deep neural network model,a kind of affinage total convolution neural network,which can quickly extract multi-scale features,is designed.Secondly,in order to solve the error problem of human body bounding frame in multi-human attitude estimation,the idea of deformation convolution is introduced into the human body attitude estimation model.And then,in order to extract human posture features efficiently,the core structure of the network is designed as a deeply separated deformed convolution kernel structure.Finally,all the structures are merged together to achieve an efficient end-to-end network.A large number of qualitative and quantitative experimental results show that the new multi-person pose estimation model proposed in this paper can still achieve higher efficiency and flexibility with higher accuracy.