Research On Human Pose Estimation Algorithm Based On Deep Learning

Human pose estimation is a technique for locating the key points of human body,such as shoulders,elbows,wrists,knees and ankles,which plays an important role in human behavior analysis.With the development of deep learning technology,human pose estimation method has made rapid progress,which usually use regression heatmap to get the location of human body's key points.In addition,when we group human key points in multi-person scenarios,the current methods do not make use of the relationship between human key points and the constraints of the human body structure.Therefore,this thesis focuses on the research of human pose estimation and its application based on deep learning.Firstly,we generally want to get a large output size while locating the key points of human body through regression heatmap by convolution neural network(CNN),but in the meanwhile large output size will lead to high computational requirements.For this reason,this thesis proposed a single-person pose estimation algorithm based on grid coding.It gets the rough location of the key points of the human body by the output probability of the grid,and then gets the precise location by calculating the output offset of the grid.Experiments show that this coarse to fine method makes the grid size much smaller than heatmap,when designing CNN architecture,the output size of CNN does not need to be enlarged by up-sampling structure,which can reduce the computational complexity of CNN.Secondly,because the number of human body is uncertain in practical application scenarios,this thesis considered how to group the detected human key points into different human bodies.Based on the previous single-person pose estimation algorithm,we propose a multi-person pose estimation algorithm based on the constraints of human body's structure.It constructs the relation between key points of human body and human Center,which predicts the center vector by CNN to point the key points of human body to the center of the relative human body.The method of grouping key points of human body analyzes the distance of human body center,which are pointed by key points of human body.When we group the key points of human body,the multi-person pose estimation algorithm is implemented.Experiments show that the proposed method can not only get the comparable results with others but also improve the detection speed.Finally,this thesis implemented an action recognition algorithm based on human pose sequence.The method first extracts human posture from each frame of video using the human pose estimation algorithm described above,and structuring human pose sequence as a specific image,and then build a CNN to classify these images to make human action recognition in video.Experiments show that the proposed method is effective and it has much higher correctness rate.