Research On Human Pose Estimation Technology Based On Deep Learning

With the rapid development of computer software and hardware technology in recent years,various human-computer interaction games and applications have been widely used.Human body posture estimation is the basis of many human-computer interaction tasks.It accurately estimates the coordinates of human joint points in the image or video information,and completes the motion recognition and behavior analysis of the characters,thus providing the next operation data for the human-computer interaction device.Therefore,it is an indispensable part of the human-computer interaction system.The human body pose estimation algorithm can be divided into the traditional graph model based method and the current deep learning based method.The traditional graph-based approach treats the human body as a series of components with strong correlations,using image structure models to simulate the appearance model of each component of the human body and the spatial constraints between components and components,and using graphical reasoning methods.Finally,optimize the position of each joint of the human body.The algorithm accuracy of this method is directly proport ional to the complexity of the algorithm.The larger the sub-model space,the more human body poses that can be simulated,and the higher the computational complexity and complexity of the algorithm.Compared with traditional methods,deep learning-based methods do not require model prior knowledge,but can achieve better results.Therefore,this thesis implements two human pose estimation networks based on deep learning.The main contributions are as follows:(1)The FPN-based two-stage R-FPN network implements the residual module and the transposed convolution on the multi-scale fusion feature obtained by the feature pyramid network.They are deep in the network layer and the upsampling multiple In a big case,good results can still be achieved.After that,through multi-stage network design and setting of multiple relay supervision points,multiple predicted thermal maps and thermal maps based on data annotation using 2D Gaussian functions are used as L2.Loss calculations to optimize n etwork parameters.The application of the above measures solves the problem that the original feature pyramid network is inaccurately positioned on the occluded joint points.(2)The Densely-hourglass Network realized by the clever design makes the hourglass module in the classic network Stacked hourglass network can be closely connected to the same channel with the same resolution and the same resolution as the Dense Net network,thus achieving less use in the network.In the case of parameters,good results can be achieved.Afterwards,the strategy of optimizing the memory usage by some concatenate and BN operations in the network achieves the purpose of compressing the memory usage,so that we can train a deeper network under limited hardware conditions.