A Comprehensive Survey of Students’ Engagement Detection for Online Classroom Management

In the quiet hum of a modern classroom, a silent shift has taken place. For decades, learning happened in physical rooms where a teacher could scan a sea of faces, reading the subtle language of furrowed brows or wandering eyes to gauge if a student was following along. Today, much of that learning has moved to screens, where the teacher is often just a small box on a monitor, and the students are scattered across the globe. In this digital space, the teacher loses the ability to see the room. They cannot tell if a student is deeply engaged, quietly confused, or simply staring blankly at the ceiling. This disconnect creates a challenge: how do you know if someone is learning when you cannot see them? The answer lies in a field of science that teaches computers to see and understand human expression, turning the webcam from a simple tool for connection into a window into a student's state of mind.

A team of researchers set out to map the current landscape of this technology, reviewing dozens of studies from the last decade to understand how machines are being taught to detect student engagement in online classes. They did not build a new system themselves; instead, they acted as cartographers, charting the territory of existing methods to see what works, what fails, and where the map remains blank. Their work focuses on computer vision, a branch of artificial intelligence that allows software to interpret visual information. In this context, the software watches a student's face through a webcam, looking for specific muscle movements around the eyes and mouth that signal emotions like happiness, confusion, or boredom. The ultimate goal is not just to label these emotions, but to translate them into a measure of engagement—determining if the student is actively participating in the lesson or drifting away.

The researchers gathered fifty-one research papers published between 2015 and 2025, filtering them through strict criteria to ensure they focused only on methods that use cameras to monitor students in real-time. They found that the process generally follows a clear path. First, the computer must find the face in the video stream. Then, it tracks that face as the student moves. Next, the system extracts specific details from the image, such as the shape of the eyebrows or the curve of a smile. Finally, a mathematical model compares these details against thousands of examples it has learned from, deciding what emotion the student is feeling and, by extension, how engaged they are.

The results of this survey reveal a technology that is surprisingly capable but still imperfect. The researchers found that many of these systems can identify student emotions with remarkable precision, with some studies reporting accuracy rates as high as 99 percent. However, these high numbers often come with a catch. The systems work best when the student is sitting still, facing the camera directly, and in good lighting. The moment a student turns their head, covers their face with a hand, or sits in a dimly lit room, the accuracy of these systems drops significantly. The researchers noted that occlusion—when part of the face is hidden—is a major hurdle. Even a simple face mask or a hand resting on a cheek can confuse the software, causing it to miss the subtle cues it needs to make a judgment.

Beyond the technical hurdles, the paper highlights a deeper complexity: the link between emotion and engagement is not a straight line. A student looking confused might be struggling to understand the material, which is a sign of active mental engagement, or they might be lost and ready to give up. A student looking happy might be enjoying the lesson, or they might be distracted by something amusing on another tab. The researchers observed that there is no universal rulebook that tells a computer exactly which emotion equals which level of engagement. Instead, most systems rely on the assumptions of the people who built them, guessing that a smile means "engaged" and a frown means "disengaged." This lack of a standard definition makes it difficult to compare different systems or to know for sure if one is truly better than another.

The review also points out that the data used to train these systems often lacks the messiness of real life. Many of the datasets consist of posed photos or videos where people are told to act out specific emotions, or they are collected from the internet where the lighting and angles are unpredictable. While these datasets help the software learn the basics, they do not always prepare it for the chaotic reality of a live online class. The researchers found that systems trained on these datasets often struggle when faced with the "wild" conditions of actual classrooms, where students move, speak, and interact in ways that are hard to predict.

Perhaps the most significant finding of the survey is the tension between accuracy and privacy. To work, these systems must constantly capture and analyze the biometric data of students—their faces, their expressions, and their movements. The researchers emphasize that this raises serious ethical questions. Storing raw video of students' faces poses a risk to their privacy, and the constant feeling of being watched could make students anxious or change how they behave. The authors suggest that the most responsible approach is to process the video data instantly on the student's own device and discard the image immediately, keeping only the final result, such as a simple score of "engaged" or "not engaged." This way, the system can help the teacher without creating a permanent record of the student's face.

Ultimately, this survey paints a picture of a technology that is powerful but still growing up. It can see faces and recognize smiles, but it is still learning to understand the human mind behind the smile. The researchers conclude that while current methods are a good start, the field needs to move beyond simple emotion recognition. Future systems must be able to handle the messy reality of occluded faces, low light, and diverse cultural expressions of emotion. They also need to find a way to balance the need for insight with the right to privacy, ensuring that the tools designed to help students learn do not end up making them feel watched or judged. The path forward is not just about making the software smarter, but about making it more human, more robust, and more respectful of the people it is meant to serve.