Research On Infant Pose Estimation Algorithms Based On Deep Learning

In recent years,with the rapid development of the field of deep learning,many fields based on deep learning have sprung up rapidly,and the pose estimation for human body is a hot research direction.It has achieved great results,such as the openpose released by Carnegie Mellon University.However,the field of action understanding and analysis of infants is still an incubation period.Based on this,this thesis firstly studies the current pose estimation model of the baby through migration learning.The main research contents are as follows:First,this thesis studies the performance of single infant pose estimation,and selects the classic network architecture stacked hourglass as the network model,and a single baby key point data set is constructed by myself.Percentage of Correct Keypoints(Normalization uses the length of the head,abbreviated as PCKh)is used to measure performance.PCK@0.5 reaches 90.8% on the test set,and the average image processing time is 0.2 seconds.Despite the high performance,in the case of a complicated plurality of babies,the posture cannot be grasped,and a misjudgment may occur.Second,in order to make the model more practical,from the perspective of multiple babies,the current star architecture openpose is selected as the network architecture,and the original caffe architecture is reconstructed with tensorflow,constructing multiple baby key point data sets in the same way as COCO.This thesis proposes a 3-stage model based on openpose,and then experiments on openpose,the trained 6-stage model,and the 3-satge model in this thesis.The 6-stage and 3-stage performances are better than openpose.For single infants,6-stage and 3-stage were 91.1% and 90.0%,respectively.In the mean average precision(mAP)comparison of multiple infants,6-stage and 3-satge were 76.8% and 75.7%,respectively.The result is basically at a level of accuracy,but with regard to runtime,the 3-stage single image is approximately 0.22 seconds,which is almost doubled compared to the 6-stage approximately 0.4 seconds.It is proved that the 3-stage adjustment is successful in the baby application scenario.This is the innovative model of this thesis,called the 3-stage infant pose estimation prediction model based on openpose.Third,the 3-stage prediction model of this thesis is applied to the video infant pose estimation,and the video is processed by cross-frame.The results show that whether the light is dim or bright,the baby's posture can be accurately positioned.The single baby video FPS value is 11,and the multi-baby video FPS value is about 8 seconds,which initially reaches the video tracking estimate.In general,this thesis takes deep learning as the background,guided by migration learning,draws on the network model structure of predecessors,and prepares the baby data set of this article.It is divided into two routes of single infant and multiple infants.Finally,the multi-baby architecture based on openpose is established as the core.The 3-stage infant posture estimation prediction model based on openpose is proposed to realize the video-based estimation of infant posture.