YOLO-APPA for student behavior detection in classrooms using adaptive boundary aggregation and feature enhancement

Abstract

Classroom behavior detection can support teaching reflection, learning engagement analysis, and classroom management by providing observable visual evidence of student activities. However, such behavioral evidence should be interpreted as an auxiliary indicator rather than a direct measurement of learning outcomes or self-regulated learning ability. In real classroom images, reliable detection is limited by small behavior regions, dense student distributions, partial occlusion, scale variation, and ambiguous boundaries between students, desks, books, phones, and background objects. Existing YOLO-based detectors provide effective one-stage detection frameworks, but they are not specifically optimized for densely seated classroom scenes in which fine-grained behaviors are small, partially occluded, and difficult to localize. To address these challenges, this study proposes YOLO-APPA, an accuracy-oriented student behavior detection framework based on YOLOv11n. The framework is designed according to the principle of preserving weak behavior-related evidence during feature extraction, feature fusion, prediction, and bounding-box regression. Specifically, YOLO-APPA integrates Adaptive Gated Selective Boundary Aggregation (AG-SBA) to enhance boundary-aware cross-level feature fusion, C3K2-PartConv to refine local feature extraction, Detect-ASFF to adaptively fuse multi-scale features, and PIoUv2 loss to improve bounding-box localization. Experiments on the Student Behaviour Detection and SCB-Dataset3-S datasets show that YOLO-APPA achieves mAP@0.5 values of 79.9% and 73.9%, respectively, and mAP@0.5:0.95 values of 56.0% and 55.5%. Compared with YOLOv11n, these results correspond to improvements of 3.8 and 3.3 percentage points in mAP@0.5, and 3.8 and 3.6 percentage points in mAP@0.5:0.95, respectively. Representative scenario-level analyses further show improved missed-target recovery in dense, occluded, far-view, and posture-similar classroom scenes. Because the complete model introduces additional computational cost relative to YOLOv11n, YOLO-APPA is positioned as an accuracy-oriented framework for GPU-assisted classroom behavior analysis rather than a lightweight edge-deployment detector. The proposed method provides visual support for more reliable classroom behavior monitoring and classroom learning-process analysis, while future work should further explore model compression, temporal behavior modeling, multimodal learning, and broader classroom contexts to improve deployment practicality and educational interpretability.