Information technologies are increasingly being integrated into all aspects of human life. Over the past few years, the use of machine learning models for human pose detection has significantly increased. As the realms of technology and physical activity converge, understanding the potential of these innovations becomes imperative for refining exercise monitoring systems. The aim of the research - evaluate the accuracy and viability of employing modern computer vision technologies in the identification of human pose during physical exercises. The study employed a combination of machine learning methods, video analysis, a review of scientific literature, and methods from mathematical statistics. The precision evaluation of contemporary machine learning models was conducted on a prepared dataset, comprising annotated images featuring students executing a body balance test with the camera positioned directly towards the subjects. The obtained data showed that both MediaPipe and OpenPose models proficiently recognize key anatomical landmarks during the conducted test. The MediaPipe model demonstrates a lower percentage of deviation from manual annotation compared to OpenPose for most key points: the mean deviation exceeds the threshold for 11 out of 15 key points and 7 out of 18 key points, as defined by the OpenPose and MediaPipe models, respectively. The most significant deviations are noticeable in the detection of points corresponding to the foot and wrist. The derived conclusions underscore the models can address only a portion of the tasks set. Essentially, this raises scepticism regarding the practical application of contemporary machine learning methods for human pose estimation without additional refinement.