Research On Human Pose Estimation Based On Deep Networks With Spatial Context

Human pose estimation is to calculate the pose parameters of human joints based on visual information.It has been widely used in the fields of medical rehabilitation,sports training,intelligent surveillance,unmanned aircraft,self-driving cars and so on.The existing end-to-end deep learning methods try to locate the joints by the convolution neural networks,but pose estimation is still a pathological problem due to the joint self-occlusion and multi-person mixing.In view of the problems,this dissertation peoposed the pose estimation method with the spatial contextual deep network from the perspective of the spatial contextual relationships between human joints.The main works of this dissertation are as follows:(1)This dissertation described the research background,significance and application scenarios of human pose estimation,and analyzed the status of pose estimation methods from the perspective of pictorial structure based models and deep learning based models.In addition,we introduced the theory of pose coding and convolution neural networks.(2)Aiming at the problem that the difficult detection joints have poor estimation effect,this dissertation proposed a human pose estimation model based on convolution neural network.The model firstly selects the credible joints according to the confidence threshold,then constructs the human pose graph,and models the spatial contextual relationships between the joints by empirical statistics.Finally,our model propagates the contextual information between the joints by convolution operations to realize the spatial inference from credible joints to incredible joints.In addition,we discussed two spatial inference ways between human joints,global spatial inference and local spatial inference.The experimental results on MPII and LSP datasets show that our method can realize the propagation of location information between the joints and could correct the incredible joints.(3)Considering that the pose estimation task depends on the multi-scale features of the images,a deep model on multi-scale spatial context is proposed.The model focuses on improving the network structure of the first stage of CPM,adds a multi-scale spatial context module,and the obtained multi-scale features include local detail features and global contextual features.Experiments on MPII and FLIC open datasets show that the proposed method can capture more effective spatial contextual features and improve the estimation accuracy when the amount of network computation is not increased much.