PAD-DETR: a student behavior detection algorithm based on knowledge distillation

Abstract Classroom behavior detection is crucial for intelligent education systems. However, real classroom environments present significant challenges, including multi-scale target detection, occlusion in densely populated settings, and small-target recognition issues. This paper proposes the pyramid adaptive detection–detection transformer (PAD-DETR), a novel algorithm for student behavior detection that enhances monitoring capabilities in intelligent classroom systems. Firstly, we develop an adaptive mixing backbone network with cross-stage dynamic mixing blocks that achieves dynamic receptive field adjustment and multi-scale feature fusion. Secondly, we introduce a context-guided adaptive pyramid network that improves the detection capability for multi-scale behaviors through rectangular perception and multi-scale context modeling. Thirdly, we design a wavelet-based contrast enhancement module specifically targeting insufficient feature contrast in classroom behavior detection. Finally, we propose PAD-DETR logic distillation combined with channel-wise distillation for feature distillation, effectively addressing insufficient small-target detection accuracy without increasing the computational load. Experimental validation on the SCB-Dataset3-S, CrowdHuman, and smart-classroom-student-behavior datasets demonstrates that compared to the baseline RT-DETR-r18 model, PAD-DETR improves the average precision (AP) by 2.7%, 2.1%, and 0.8%, while enhancing the AP 50 by 3.6%, 1.4%, and 0.9%, respectively. These results confirm that PAD-DETR effectively addresses multi-scale challenges in classroom behavior detection, providing an efficient solution for smart classroom behavior analysis.