Student Classroom Behavior Recognition Based on Improved YOLOv8s

1. Problem Statement

The paper addresses the challenges of automatically recognizing student behaviors in real-world classroom environments using computer vision. While deep learning has advanced this field, existing models struggle with three specific characteristics of classroom scenes:

• Dense and Small Targets: Students are often densely packed, and many appear as small objects in the frame, leading to missed detections.
• Severe Occlusion and Similarity: Frequent mutual occlusion (e.g., students sitting behind desks) and high visual similarity between fine-grained behaviors (e.g., "sitting upright" vs. "writing" vs. "reading") cause false positives and low discrimination.
• Class Imbalance: Training data often suffers from long-tail distributions where common behaviors (e.g., listening) have abundant samples, while rare behaviors (e.g., hand-raising, standing) are underrepresented, causing models to bias toward majority classes.

2. Methodology: ALC-YOLOv8s

The authors propose ALC-YOLOv8s, an improved version of the YOLOv8s (You Only Look Once version 8 small) architecture. The model retains the lightweight nature of the baseline while introducing three specific modules to address the aforementioned challenges:

A. SPPF-LSKA (Receptive Field & Context Enhancement)

• Location: Replaces the standard SPPF (Spatial Pyramid Pooling - Fast) module in the backbone.
• Mechanism: Integrates LSKA (Large Kernel Separable Attention) into the pooling structure.
• Function: While standard pooling aggregates global information, it often loses structural details. LSKA captures dependencies between spatial and channel domains using large kernels. This enhances the model's ability to understand the spatial context surrounding a student, helping to distinguish occluded or small targets by leveraging surrounding structural cues.

B. CFC-CRB and SFC-G2 (Feature Fusion Optimization)

• Location: Integrated into the backbone and neck networks to improve cross-layer interaction.
• CFC-CRB (Context Feature Calibration Block): Uses cascaded pyramid pooling to generate multi-scale context. It adaptively reweights spatial features based on global priors, ensuring that deep semantic information is consistent with local details.
• SFC-G2 (Spatial Feature Calibration G2): A module that aligns high-level (semantic) and low-level (detail) features. It uses lightweight convolutions to predict spatial displacements, resampling features to correct offsets. A gating mechanism then adaptively fuses these streams.
• Function: These modules solve the contradiction where deep layers have strong semantics but weak details, and shallow layers have rich details but weak semantics. They ensure fine-grained behavioral cues (like arm position or head orientation) are preserved during fusion.

C. ATFLoss (Loss Function Optimization)

• Location: Replaces the standard classification loss function.
• Mechanism: Implements Adaptive Threshold Focal Loss (ATFLoss).
• Function: This loss function applies differentiated weights based on sample difficulty and class frequency. It suppresses the gradient contribution of "easy" and "majority" samples while amplifying the learning signal for "hard" and "minority" classes (e.g., hand-raising). This mitigates class imbalance and forces the model to focus on difficult decision boundaries.

3. Key Contributions

1. Architecture Improvement: Proposes a tailored YOLOv8s variant (ALC-YOLOv8s) specifically optimized for the dense, occluded, and imbalanced nature of classroom scenes.
2. Module Innovation: Introduces a novel combination of SPPF-LSKA for context, CFC-CRB/SFC-G2 for precise feature fusion, and ATFLoss for balanced learning.
3. Comprehensive Validation: Conducts extensive ablation studies and comparative experiments against mainstream detectors (SSD, Faster R-CNN, RT-DETR, and various YOLO versions) on a self-constructed dataset.

4. Experimental Results

The experiments were conducted on a dataset containing seven behavior categories: sitting upright, looking down, looking around, reading, writing, standing, and hand raising.

• Ablation Study:

  • The baseline YOLOv8s achieved an mAP50 of 0.861 and mAP50-95 of 0.715.
  • The full model (ALC-YOLOv8s) achieved mAP50 of 0.879 (+1.8%) and mAP50-95 of 0.736 (+2.1%).
  • The combination of feature fusion enhancements (B) and loss optimization (C) showed the strongest synergistic effect before adding the context module.

• Comparative Performance:

  • ALC-YOLOv8s outperformed all compared models, including YOLOv5s, YOLOv11s, and the two-stage Faster R-CNN.
  • Notably, while Faster R-CNN had a decent mAP50 (0.822), its mAP50-95 was significantly lower (0.577), indicating poor performance on high-IoU (tight bounding box) localization, which ALC-YOLOv8s handled superiorly.
  • The proposed model achieved the best trade-off between Precision (0.826) and Recall (0.831).

5. Significance and Conclusion

• Practical Application: The study demonstrates that the improved model is robust enough for real-time, automatic analysis of student engagement in complex, crowded classrooms.
• Technical Insight: It validates that addressing contextual modeling (via LSKA), feature alignment (via SFC-G2), and sample imbalance (via ATFLoss) simultaneously is crucial for fine-grained action recognition in educational settings.
• Future Work: The authors note that while the current model is effective, future research should focus on expanding the dataset scale, increasing scene diversity, and incorporating temporal information (video sequences) to further improve generalization across different teaching scenarios.