Deformable Pyramid Sparse Transformer for Semi-Supervised Driver Distraction Detection.

Area of Science:

• Intelligent Transportation Systems
• Computer Vision
• Machine Learning

Background:

• Driver attention is crucial for intelligent transportation safety.
• Manual annotation for driver distraction detection models is costly and time-consuming.
• Existing models struggle with limited labeled data.

Purpose of the Study:

• Propose an adaptive semi-supervised driver distraction detection framework.
• Improve model performance with limited labeled data.
• Enhance real-world driver monitoring systems.

Main Methods:

• Utilizes teacher-student learning and deformable pyramid feature fusion.
• Incorporates an adaptive pseudo-label optimization strategy with category-aware thresholding and confidence weighting.
• Integrates a Deformable Pyramid Sparse Transformer (DPST) module into a YOLOv11 detector.
• Employs teacher-guided feature consistency distillation.

Main Results:

• Achieves robust and scalable distraction detection using limited labeled and abundant unlabeled data.
• The DPST module enhances fine-grained perception of subtle driver behaviors.
• Outperforms fully supervised baselines on the Roboflow Distracted Driving Dataset in mAP metrics.
• Demonstrates a balanced trade-off between precision and recall.

Conclusions:

• The proposed framework offers an effective solution for driver distraction detection under limited annotation conditions.
• Enables practical and scalable driver monitoring systems.
• Mitigates the impact of noisy pseudo-labels through feature consistency distillation.