PERFORMANCE EVALUATION OF RECENT YOLO VERSIONS FOR CLASSROOM STUDENT BEHAVIOR DETECTION

Abstract

The increasing adoption of smart classroom systems underscores the need for automated, objective, and real-time  monitoring of student behavior to support effective teaching and learning. Computer vision–based object detection, particularly the You Only Look Once (YOLO) family, has shown strong potential for this task. However, existing studies predominantly evaluate YOLO models in isolation or across different frameworks, resulting in biased comparisons. To address this gap, this study presents a controlled intra-family comparative evaluation of four recent YOLO generations YOLOv8, YOLOv10, YOLOv11, and YOLOv12 across three weight variants (nano, small, and medium), yielding 12 model configurations. All experiments were conducted under a uniform training pipeline and computing environment using an NVIDIA T4 GPU to ensure fair benchmarking. Model performance was assessed using Precision, Recall, F1-Score, mean Average Precision (mAP), inference speed (FPS), and computational complexity. The results reveal a consistent trade-off between detection accuracy and inference speed: YOLOv12m achieves the highest detection accuracy but the lowest FPS due to increased architectural complexity. At the same time, YOLOv10n offers the fastest inference at the cost of reduced reliability for subtle behaviors. Within the scope of the evaluated dataset and controlled classroom setting, YOLOv8s and YOLOv11s demonstrate the most balanced accuracy–speed performance, making them suitable candidates for real-time  classroom monitoring under similar conditions. This study provides practical insights for researchers and developers by offering an objective benchmark and model-selection guidance tailored to smart classroom applications, while accounting for dataset and environmental constraints.