Real-Time Detection of Student Classroom Behavior Patterns Using the YOLO11-GELSK Multi-Scale Fusion Architecture

Detecting student classroom behaviors from ordinary classroom videos is complicated by weak features for small targets such as rear-row mobile phones, feature loss caused by inter-student occlusion, imbalance between frequent learning behaviors and rarer off-task behaviors, and accuracy degradation when a detector trained in one classroom setting is transferred to another. Existing YOLO-series detectors are typically limited by shallow detail extraction and insufficient multi-scale fusion, making it difficult to reconcile accuracy with real-time inference. To address these limitations, we propose YOLO11-GELSK, an optimized detection architecture built upon YOLO11s that integrates a gated Global Edge Information Transfer (GEIT) module with a star-shaped Large-Kernel Separable Attention (LSKA) fusion neck. The gated GEIT module extends the shallow, Sobel-based edge-transfer design of SCB-YOLO with an image-level channel gate and cross-stage residual transfer, strengthening contour extraction for fine-grained actions such as bowing the head, leaning over the table, or holding a phone, while the star-fusion neck combined with LSKA attention captures global spatial context, mitigating occlusion and missed detections of distant targets. A scale-adaptive Focal-CIoU composite loss balances training across behavior categories to suppress false positives and negatives for underrepresented classes, and a Maximum Mean Discrepancy (MMD) unsupervised domain-adaptation term aligns neck-level features across classroom subsets; L2 weight decay and stochastic depth further curb overfitting. Six public SCB-Dataset3 behavior categories—hand-raising, reading, writing, using a phone, bowing the head, and leaning over the table—are modeled as an object-detection task, and a multi-target probability fusion mechanism aggregates per-student outputs into a continuous, frame-level off-task score. On a recording-grouped validation split of SCB-Dataset3, YOLO11-GELSK attains 91.7% mAP@0.5 and 90.2% macro-F1, exceeding YOLO11s, ResNet50-based Faster R-CNN, and EfficientNet-B3-based RetinaNet baselines by 5.2–11.9 mAP points and the closest classroom-specific detector, an s-scale re-implementation of SCB-YOLO, by 2.5 points; recording-grouped five-fold stratified cross-validation further yields 91.5 ± 0.4% mAP@0.5 and 89.8 ± 0.4% macro-F1 (the paired per-fold gain over YOLO11s is positive on every fold, 5.0–5.5 mAP points), while the complete video pipeline sustains 33.6 FPS, providing a lightweight and accurate solution for intelligent classroom behavior monitoring.