DETERMINATION AND COMPARISON METHODS OF BODY POSITIONS ON STREAM VIDEO

Abstract

Context. One of the tasks of computer vision is the task of determining the human body in the image. There are many methods to solve this problem, some are based on specific equipment (motion capture, kinect) and provide the highest accuracy, some give less accuracy but do not require additional equipment and use less computing power. But usually, such equipment has a high cost, so to ensure the low cost of developments designed to determine the body in the image, you should develop algorithms based on computer vision technology. These algorithms can then be applied to various fields to analyze and compare body positions for a variety of purposes.
 Objective. The aim of the work is to study the effectiveness of existing libraries to determine the human body position in the image, as well as methods for comparing the obtained poses in terms of speed and accuracy of determination.
 Methods. A set of libraries and pose comparison algorithms were analyzed for the purpose of developing a system for determining the correctness of exercise by the user in real time. OpenPose, PoseNet and BlazePose libraries were analyzed for their suitability in recognizing and tracking body parts and movements in real-time video streams. The advantages and disadvantages of each library were evaluated based on their performance, accuracy, and computational efficiency. Additionally, different pose comparison algorithms were analyzed. The effectiveness of each algorithm was evaluated based on their ability to accurately determine and compare body positions.
 As a result, the combination of BlazePose and weighted distance method can achieve the best performance in pose recognition, with high accuracy and robustness across a range of challenging scenarios. The weighted distance method can be further enhanced with techniques such as L2 normalization and pose alignment to improve its accuracy and generalization. Overall, the combination of the BlazePose library and weighted distance methods offers a powerful and effective solution for pose recognition, with high F1 index.
 Results. Existing models for determining poses have shown similar results in the quality of determination with a run-up of about 2%. When developing a cross-platform software product, the BlazePose library, which has an API for working directly in the browser and on mobile platforms, has a significant advantage in speed and accuracy. Also, as the library uses extended 33 keypoint topology it becomes applicable to a wider list of tasks. In the study of comparison methods, the greatest influence on the results was exerted by the quality of pose determination.
 Conclusions. Among the methods of comparison, the method of weighted distances showed the best results. The speed of position determination is inversely proportional to the quality of determination and significantly exceeds the recommended value – 40ms.