| The human pose estimation has limitations: the uncertainty and complexity of different appearances,genders,clothing,occlusion,background and lighting conditions.How to effectively deal with the influence of different conditions and an effective method is the focus of this thesis.Firstly,for the human body is difficult to detect parts,using the improved multi-scale feature and multiple dimension convolution kernels,this thesis proposed a coordinate transformation network based on the feature pyramid,which firstly obtains the multi-scale features of one picture to solve the environmental interference problem,then increases the convolution kernel of the local area of the network for the parts with severe occlusion,and processes the losses of hard samples;finally,the method converts the keypoints into the gaussian heatmap and changes the heatmap into coordinates,so as to realize the end-to-end training.Experiment show that the improved features are helpful to identify the difficult parts,and the network computation is less.Secondly,to find the non-euclidean spatial structure relation between the keypoints,this thesis proposed a method of human pose estimation based on graph convolutional neural network,which improves the accuracy of human pose estimation.Its input data structure for the space and time graph,the keypoints for the nodes and connection between the bone for the edges to make up the figure of a single frame structure.In this method,the localization of a node is based on the spatiotemporal graph structure after affine transformation of the position of the node,and then the region where the node is located is transformed into a localization node process through a graph neural network.The experiments show that the network can improve the detection performance.The human pose estimation method has the ability of detecting human body parts and joints.Several groups of experiments were carried out for mpii test set and coco test set,and Pckh and AP are used in the experiments. |
PDF Full Text Request