Human physical exercise recognition and analysis from video streams is a relevant scientific and applied problem in modern sports, physical education, and rehabilitation. It enables the automated assessment of movement quality, increases the objectivity of training control, and supports the timely correction of technique. In contrast to purely expert observation, videobased analytics reduces the influence of subjective judgments and creates prerequisites for designing personalized training programs that account for an athlete’s individual motor patterns. The aim of this paper is to review modern models and methods for human physical exercise recognition and to generalize key application directions for improving training effectiveness and safety, with an emphasis on markerless and scalable solutions suitable for practical use. The review analyzes recent research publications devoted to movement recognition and the automatic evaluation of exercise execution techniques, as well as approaches to feature engineering, model training, and validation protocols. The main classes of solutions used in practice are summarized: (1) biomechanical and physiological approaches based on specialized hardware–software systems; (2) sensor-based technologies relying on wearable devices (inertial sensors, GPS/IMU), which provide quantitative workload indicators; and (3) computer vision and machine learning methods, including deep learning architectures for pose estimation, object tracking, and action classification in video. The analysis indicates that the most promising directions are hybrid solutions that combine spatial feature extraction with temporal modeling of movement dynamics, support real-time feedback, and allow task-specific fine-tuning for various exercise types. Typical limitations of video-based approaches are also highlighted, including sensitivity to acquisition conditions, camera viewpoints, occlusions, illumination changes, and video quality, as well as the lack of standardized validation procedures. Video analytics combined with artificial intelligence provides a foundation for more accurate, scalable, and personalized control of exercise techniques and for injury prevention. Future research should focus on integrating multimodal data (video plus wearable sensors), developing unified evaluation protocols, and improving the interpretability of model outputs for coaches and rehabilitation professionals.