Researches On Human Pose Estimation Based On Deep Learning

The human pose estimation task has been of great interest during the development of computer vision.The pose estimation task is also one of the challenging tasks in industry and academia,where the goal is to make the machine detect key nodes in human samples,such as nose,left and right shoulder,ankle,and wrist,as much as possible.With the introduction and application of deep neural networks,the human pose estimation task is often a fundamental research task for predicting human behavior patterns,providing basic predictive capabilities for tasks such as pedestrian detection,sample reidentification,special behavior detection,and human-computer interaction.The current mainstream human pose estimation network framework is divided into two application-based networks,namely human structure application network and key point detection network.In this thesis,the human pose and estimation tasks and the mainstream pose estimation networks are carefully profiled,and a new network search space is designed by combining the microscopic network architecture search method,from which a high-performance keypoint encoder and an adaptive keypoint decoder are searched.At the same time,this thesis presents a generalized study of the searched keypoint detection network in terms of attention mechanism and multidimensional scaling.The main work of this thesis includes:(1)Lightweight network architecture search.Microscopic network architecture search algorithms have many drawbacks,two of which are large memory consumption and performance collapse.In this thesis,we propose to perform channel sampling or binarized network processing on the hypernetwork,which enables to perform network architecture search directly on large datasets,but also leads to increased performance collapse.In this thesis,we add an unbiased noise with small variance,and channel weight some of the features learned by the operator,which can alleviate the problem of performance collapse under local networks,improve the stability of network architecture search,and make the searched network have good performance.(2)Multi-scale based key point detection network architecture search.In this thesis,on the basis of keeping the high-resolution features,we start from the scale-sensitive perceptual field change module,and search the multi-scale feature fusion network from the perspective of network autonomy in choosing the operator connection method,so that the representations of different resolutions can promote each other to improve,and finally output the pose estimation heat map with high localization accuracy.Second,this thesis conducts a generalized study on the spatial and channel attention mechanisms and multidimensional model scaling on the searched key point detection network.This thesis designs and implements a multi-scale network architecture search method more suitable for pose estimation,which is able to perform fast network architecture search.The searched networks are fully experimented on public datasets MPII and COCO,and good experimental results are obtained.