Student Classroom Behavior Recognition Based on Improved YOLOv8s

Researchers Xiang Gao and Shuai Hang from an undisclosed institution have published a paper on arXiv proposing ALC-YOLOv8s, an enhanced version of YOLOv8s designed to recognize student behaviors in real classroom settings. The model addresses common challenges like dense student targets, numerous small objects, frequent occlusions, and imbalanced class distributions by integrating three key architectural improvements. First, SPPF-LSKA (Large Separable Kernel Attention) replaces the original SPPF to extract richer contextual features. Second, CFC-CRB (Cross-Feature Convolution with Channel Refinement Block) and SFC-G2 (Semantic Fusion Convolution with Group 2) modules optimize multi-scale feature fusion, critical for detecting both small hand raises and full-body postures. Finally, ATFLoss (Adaptive Threshold Focal Loss) improves learning for minority classes (e.g., sleeping) and hard examples.

Experimental results on a custom classroom dataset show that ALC-YOLOv8s outperforms the baseline YOLOv8s by 1.8% in mAP50 and 2.1% in mAP50-95. Comparisons with other mainstream detection methods (including Faster R-CNN and YOLOv7) confirm its suitability for complex, real-world classroom scenarios. The paper provides detailed ablation studies validating each component's contribution and visualizations of detection results on crowded images. While the work is preliminary and dataset specifics remain limited, it demonstrates practical advances in applying computer vision to educational environments, potentially enabling automated teaching quality analysis and real-time student engagement monitoring.