Performance Evaluation System Based on Classroom Scene Analysis Using a Hybrid Deep Learning Technique

Purpose: To create a hybrid model for in-class video analysis that can recognise and interpret student behaviour, participation, and interaction to enable data-informed classroom management. Design/Methodology/Approach: Model development with experimental evaluation on a supervised and annotated video dataset. The research population and context include undergraduate students and instructors from the University of Technology, Iraq, as well as public classroom footage from YouTube. A purposive sampling strategy covered lectures, laboratories, and seminars; the final sample comprises 120 frames. Data were preprocessed (resizing and normalisation) and manually annotated (bounding boxes and behaviour labels). YOLOv8 performed object detection, CNNs extracted discriminative features, LSTMs modelled temporal dependencies for behavioural prediction, the pipeline was trained end-to-end and evaluated on a held-out test set. Findings: The YOLOv8–CNN–LSTM architecture achieves 97% accuracy, demonstrates stable training/validation curves, and outperforms previous baselines on similar tasks. Research Limitations: The study used specific classroom settings and stationary cameras with limited perspectives; a multi-institutional dataset would be needed to generalise the results for broader use. Practical implication: Provides information-actionable representations of student engagement and engagement in ways that promote a timely intervention and personalised instruction. Social implication: Provides the support and encouragement for inclusive learning environments concerning student learning outcomes and wellbeing. Originality/value: An integrated end-to-end hybrid pipeline specific to classroom scenes, which moves from static recognition to predictive behaviour analysis of intelligent performance evaluation.