Research And Implementation Of Infants Abnormal Behavior Based On Human Pose Estimation

General movements are widely used in clinical practice,for infants born between two and six months old,observing the whole body of babies with a non-contact form to infer the infant’s neural development.However,this evaluation method is subjective and takes a long time for the observer to train.Current this method progress slow in China.With the development of deep learning,general movements based on computer vision is constantly improving,and there is huge room for development in the newborn field.Based on the research of domestic and foreign 3D pose estimation and motion feature analysis,using general movements as the basis for infant motion feature analysis.Based on human pose estimation to make motion feature extraction and analysis from newborn’s whole body motion video.Doing this to judge the infant’s neural development state.The main work of this thesis as follows:(1)Based on existing human key point pose estimation method based on generating heatmaps.Using Res Net and deconvolution layer to make a two-dimensional human key point reorganization model.After testing,it has a good effect on infant detection,and the recognition accuracy has been improved from 85% to 86.9%.(2)For the recognition of three-dimensional key points of the human body,the deep learning method is used to upgrade the two-dimensional coordinates of the human body.Based on Dense Net,with one-dimensional convolutional neural network is used to design a human body’s two-dimensional key point coordinate ascending model.After the experiment and comparing with the original model,the accuracy of the three-dimensional key point coordinators estimation has improved by about 2%.(3)According to the key points of the three-dimensional pose coordinates of the human body,a set of angle extraction methods for the key points in the three-dimensional space of the human body based on quaternion are proposed,and the principal component analysis method is used to classify and detect the extracted parameters.And try to use principal component analysis to reduce the dimensions of multi-dimensional features and improve the detection performance of the model while ensuring the recognition accuracy.Combining all the above research results,design an infant abnormal behavior detection system,uploading data via the embedded system,using baby motion video to verify and test the algorithm,after testing,the expected performance has been archived.