Human Pose Estimation Based On Convolutional Neural Network

Human pose estimation is an important basis for image analysis and behavior recognition.It can be used to assist the understanding of image content by detecting the position of human joint points in the image.Human pose estimation has broad application prospects,and is increasingly used in intelligent monitoring,human-computer interaction and motion analysis.Due to the problems of occlusion,multi-view changes and background interference,traditional algorithms have been unable to achieve high accuracy,so it is of great significance to carry out research on it.With the arrival of the era of big data and the improvement of hardware computing ability,convolutional neural network has re-entered the public's vision,and has achieved amazing results in various areas of computer vision,such as face recognition,object detection and so on.More and more people begin to try to use convolutional neural network to solve problems that traditional methods can not solve.In this paper,a convolutional neural network is chosen to solve the problem of human pose estimation.Different from traditional methods,convolutional neural network has powerful learning ability,which can effectively utilize context information of joint points,so as to make more accurate reasoning and prediction,and better solve the traditional problems such as occlusion,multi-view changes and so on.The main research work in this paper is as follows:(1)The shortcomings of DeepPose method in prediction accuracy and network model size are analyzed.A global attitude-constrained network structure is designed by stacking multiple convolutional layers containing 3*3 smal convolution kernels to replace of the convolution layer containing larger convolution kernels in DeepPose structure.This operation increase the depth of the network and the capability of non-linear mapping,and make the improvement of the accuracy of joint detection;and use the global average pooling layer instead of the fully connected layer,which greatly reduce the parameter size of the network model,makes the network model occupy less memory resources and is easy to migrate to some resource-constrained embedded devices.(2)For the problem that the deep learning model is easy to over-fit in the case of small sample set,an objective function including the spatial constraint relationship between human joints is designed.Shape context algorithm is used to measure the shape similarity between the predicted value set and the real value set of all nodes.This similarity includes the spatial constraints between the joint points.The objective function is composed of the penalty term and the square loss function,so that the network can implicitly learn the spatial distribution of human joint points and reduce the degree of over-fitting of the network.(3)The pose estimation method of this paper is applied to fall detection.A part of image data simulating the falls of the human body is collected.Joint posture map is constructed based on the results of pose estimation.Use the LeNet-5 network to train and test on this data set.The test results are analyzed by the accuracy,recall and ROC curves,and the effectiveness of this method for fall detection is verified.