YOLO11n-LDS: A Lossless Downsampling and Direction-Aware Framework for Robust Driver Distraction and Fatigue Detection

Engineering notes

Experimental results demonstrate that YOLO11n-LDS outperforms the baseline YOLO11n model, improving Precision, F1-score, mAP@50, and mAP@50:95 by 2.51%, 1.22%, 1.43%, and 0.78%, respectively. It also achieves superior performance compared with several state-of-the-art object detection models.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract Driver distraction and fatigue are major factors contributing to traffic accidents; however, accurate recognition of subtle driver behaviors remains challenging due to small-scale visual cues, complex in-vehicle backgrounds, and significant variations in illumination and viewpoint conditions. To address these challenges, this paper proposes YOLO11n-LDS, a driver monitoring framework that establishes a unified feature enhancement paradigm for fine-grained behavior recognition. The proposed framework integrates three complementary mechanisms: Lossless Downsampling Convolution (LDSConv) for preserving fine-grained spatial information during downsampling, Directional Feature Aggregation (DFA) for enhancing orientation-sensitive contextual representations, and Spatial–Channel Interaction Attention (SCIA) for improving semantic discrimination through joint spatial–channel modeling. Through the synergistic interaction of feature preservation, directional perception, and semantic refinement, YOLO11n-LDS generates more discriminative feature representations for driver distraction and fatigue detection. To evaluate the proposed method, experiments are conducted on the self-constructed SFHDCL dataset, which contains RGB and infrared (IR) data collected under diverse driving conditions, including daytime, nighttime, and rainy weather. Experimental results demonstrate that YOLO11n-LDS outperforms the baseline YOLO11n model, improving Precision, F1-score, mAP@50, and mAP@50:95 by 2.51%, 1.22%, 1.43%, and 0.78%, respectively. It also achieves superior performance compared with several state-of-the-art object detection models. A prototype driver monitoring system was further developed to validate its practical applicability. The results demonstrate that the proposed method provides accurate and reliable driver behavior recognition in complex in-vehicle environments, showing strong potential for intelligent driving safety applications.

Links and sources

• PDF from original source

Comments