Research On Human Pose Estimation Based On Static Image

With the widespread popularity of electronic devices,people are looking forward to a more intelligent and convenient way of interaction,so the issue of human pose estimation has always been a research hotspot.This thesis will focus on the problems of human pose estimation methods and propose some methods to address these problems.Human pose estimation methods can be roughly divided into model-free and modelbased methods.First,the performance of model-free algorithms often depends on the size of training set,so this type of methods often fail to get good results under the sparse training set.Secondly,although the methods based on convolution neural network can often obtain good prediction results,the network models generated by such methods are often relatively large,which can't transplante into embedded devices.At the same time,it also takes a lot of time to train and test these methods.Finally,in the process of resolving human pose with model-based methods,it is often necessary to use feature pyramid to deal with the different scales object,which will cause a lot of unnecessary calculations.This thesis proposes some methods to address above problems.The contributions are summarized as follows.(1)Applying the Local Subspace method to the human pose estimation task,it can still obtain good prediction results under the sparse training set,which overcomes the problem that the model-free learning method is very sensitive to the size of training set.Then we directly get the closed-form solution of the relevant parameters by reorganizing the error function,which can greatly reduce the time spent in the training process,allowing us to quickly set the most suitable number of subspace and neighbor subspace.(2)Applying a new module to replace the convolution operating with large kernel size in the original network structure generated by Convolution Pose Machine method.The new module has the ability to use a limited number of parameters to multi-scale features in parallel.More importantly,these multi-scale features can really help us to effectively predict the position of human part,which achieves the purpose of compressing model.(3)Analyze the main time-consuming factors of each operation in the process of resolving human pose based on the Mixture of Parts model method,and then we propose a method based on important parts to accelerate the resolving process.This method utilizes the strong correlation between human parts to reduce the number of resolutions in the feature pyramid through a few important parts,so that the operations that originally needed to be performed in all resolutions now need only be performed at a small number of effective resolutions.