A WAD-YOLOv8-based method for classroom student behavior detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional object detection models face limitations in complex environments like classrooms, struggling with restricted receptive fields and insufficient multi-scale feature learning.
• Fixed convolutional kernels in existing backbones hinder the ability to capture diverse features crucial for accurate behavior analysis.

Purpose of the Study:

• To enhance the YOLOv8 model for improved complex classroom behavior detection.
• To overcome limitations in receptive field and multi-scale feature learning inherent in standard object detection architectures.

Main Methods:

• Introduction of a novel Convolutional Attention-based Cross-Stage Partial Network (CA-C2f) module for comprehensive receptive field fusion and adjustment.
• Integration of an attention-based 2D Position Encoding-Multi-Head Attention (2DPE-MHA) module to capture long-range dependencies.
• Incorporation of a dynamic sampling factor (Dysample) to focus on detailed regions and prevent information loss.

Main Results:

• The enhanced YOLOv8 model demonstrated superior performance on benchmark datasets (SCB, SCB2, SCB-S, SCB-U) compared to existing methods.
• Achieved significant improvements in mean Average Precision (mAP@0.5) ranging from 2.2% to 18.7% and mAP@0.5:0.95 from 2.3% to 14.8%.
• Maintained real-time inference speeds, outperforming other object detection models.

Conclusions:

• The proposed model effectively addresses challenges in complex classroom behavior detection, particularly for multi-scale, occluded, and small targets.
• The integration of CA-C2f, 2DPE-MHA, and Dysample modules significantly boosts detection accuracy and detail preservation.
• This model offers a practical solution for real-time classroom behavior monitoring, assisting educators and administrators.