Using a neural network in skeleton-based multiscale human pose estimation
A.R. Akhatov, I.Q. Himmatov, Arvind Dagur · 2024
Evidence shows that in recent years the task of assessing a human&s;s condition from a photo or video taken from a camera has been improved, it allows you to form a human&s;s pose in 2D and 3D and identify a human&s;s movement by dividing it into joints and points. One of the reasons for this trend is such applications are increasing: computer, interaction with robots, virtual films and video games; analysis of the actions and activities of athletes; video surveillance, security control, etc. Another reason is the successful application of neural network architecture to computer vision problems. This is an signification step towards recognizing persons in videos and images, as well as psychological identification of characters. Nowadays, the identification of a human by his actions is one of the processes that are difficult to set up and master compared to other identification technologies. Therefore, in this study, the recognition of human poses and the construction of graphs using neural networks and other tools are considered as initial tasks. Human pose estimation is a special case of the image segmentation problem in the computer vision department, which consists in detecting the movements from parts of the human body of images or videos (considered as a sequence of images). Often a human&s;s position is replete with associated key points that correspond to the joints (shoulders, elbows, arms, hips, knees, feet) and other key points (neck, head). This task can be considered in two or three dimensions, which determines the complexity of the task and the practical application of the results ( Akhatov et al. 2021 ). The task can also be divided into two subtypes: Single Human Pose Estimation, Multi Human Pose Estimation.