YOLO-MIRNet: detection of distracted driving with enhanced generalization using a lightweight multi-scale interaction...

Distracted driving is one of the key factors contributing to road crashes. Real-time and robust monitoring of drivers' states is of great significance for improving the level of road traffic safety. Existing methods based on convolutional neural networks still face challenges in terms of generalization ability and practicality under cross-scenario, cross-device, and complex environmental conditions.

To enhance the adaptability and robustness of the model in diverse real driving scenarios, this paper proposes a lightweight detection framework with enhanced generalization ability, namely YOLO-MIRNet. Based on the YOLOv8 architecture, the model integrates three core innovations: first, a Global and Detail Feature Aggregation (GDFA) module is designed, which collaboratively extracts multi-scale features through multi-branch dilated convolutions and channel shuffling; second, a Multi-scale Adaptive Gated Fusion (MSAGF) module is constructed, which realizes adaptive fusion of cross-scale features by utilizing pixel-wise attention and gating mechanism; third, an improved Lightweight Channel Attention (iMLCA) module is embedded to strengthen the discriminative ability of key features with low computational cost. Model training and evaluation are conducted on the well-partitioned public StateFarm dataset and the more challenging AUC dataset (partitioned by driver).

Experimental results demonstrate that YOLO-MIRNet achieves a detection precision of 99.92% and a mean Average Precision (mAP) of 99.48% on the StateFarm dataset; it also attains a precision of 87.03% and an mAP of 93.99% on the AUC dataset, significantly outperforming the comparison models. With a parameter count of only 12.68M, the model exhibits excellent lightweight characteristics, and the model's single-frame inference time remains stable between 42 and 57 milliseconds. The generalization performance test results are prominent: in the cross-dataset evaluation ("StateFarm→AUC"), the precision reaches 67.03%; under the reverse setting ("AUC→StateFarm"), the accuracy is further improved to 72.61%. In addition, on the 100-driver multi-modal dataset containing day and night scenarios, the model's detection performance during both day and night significantly surpasses the baselines, demonstrating strong environmental adaptability.

The proposed YOLO-MIRNet framework achieves an optimal balance among detection accuracy, model lightweight, and cross-domain generalization ability. It provides an efficient and reliable technical solution for real-time monitoring of distracted driving, which is conducive to promoting the practical application and deployment of intelligent traffic safety systems.