OVERVIEW OF TECHNOLOGIES FOR DETERMINING THE POSITION OF THE HUMAN BODY

Abstract

Background. The assessment of the position of the human body is an important task of computer vision, which includes the prediction of 3D and 2D coordinates, the location of the joints of the human body based on images and videos. Materials and methods. COCO, MPII Human Pose and Human3.6M datasets, MPJPE, mAP and PCK metrics for evaluating results, as well as deep neural networks for training models were used. Results. The article presents a comparison of the training results of models on several datasets, including COCO, MPII Human Pose and Human3.6M, based on MPJPE, mAP and PCK metrics. In addition, the advantages and disadvantages of various methods and models, as well as their applicability to various problems, are discussed. Conclusions. Modern methods based on deep neural networks demonstrate high accuracy and efficiency in solving the problem of estimating body position. However, such models require large computational resources and training time. The choice of a specific model and method depends on the requirements of a particular task and the available computing resources.

Keywords

• human pose estimation
• deep learning
• convolutional neural networks
• recurrent neural networks
• bottom-up approach
• top-down approach
• datasets
• evaluation metrics
• challenges