This review systematizes research on automatic detection of student engagement in online learning using computer vision and deep learning from 2020 to 2025. Engagement is considered a multidimensional construct that includes emotional, cognitive, and behavioral components that play an important role in academic performance. The review analyzes the datasets used (public and specialized), feature extraction methods (facial expression, gaze direction, head posture, body movements), model architectures (CNN, LSTM, transformers), and approaches to multimodal integration. Although the transition from experimental solutions to complex real-time systems is demonstrated, persistent difficulties in model generalization, lack of diverse data, and ethical risks related to privacy and processing of personal information remain. It concludes that there is a need to develop standardized, ethically sound, and adaptable methods for assessing participation