Vision-Based Solutions for Human Activity Recognition across Course and Fine Levels

This paper focuses on detection of human activities both at course level and finer levels. A course level activity like squats and finer level activity like hand gestures were detected using computer vision techniques and machine learning algorithms. The study integrates state-of-the-art technologies such as Media Pipe to detect pose landmarks and recognize hand gestures in real time. In squat analysis, an angle between pose landmark and real time pose was determined that enabled assessment of correctness of the squats. While for hand gesture recognition, a custom model training on a dataset with noisy backgrounds was done and validation accuracy of 70% was obtained. Despite the lower initial accuracy, the approach ensures improved robustness in practical applications. As a case study, the gesture recognition system was deployed for applications including scrolling of the screen, changing the slides of presentation.