Driver distraction detection method based on YOLOv11s

Abstract

Distracted driving is a key factor leading to traffic accidents. Timely identification and intervention of this behavior are of great significance for improving traffic safety. To this end, this paper proposes a driver distraction detection method based on the improved YOLOv11s. This method introduces an efficient multi-scale attention (EMA) mechanism in the neck network, enhancing the feature extraction capability through multi-scale feature extraction and cross-space learning architecture. To further improve detection accuracy, Inner-WIoU is adopted as the bounding box regression loss function. Its dynamic weight allocation strategy can assign higher loss weights to difficult-to-identify samples such as occlusion or abnormal postures, guiding the model to focus on these samples and effectively alleviating the problem of sample imbalance. Experimental results on the SFDDD distracted driving dataset show that the improved YOLOv11s algorithm significantly improves detection accuracy.