DdbiNet: A dangerous driving behavior identification network based on Resnet18

The direct correlation between drivers' dangerous behaviors and vehicle safety underscores the critical need for timely identification of such behaviors to enhance traffic safety. To enhance the detection of dangerous driving behaviors, we incorporated the advanced SCConv module into the ResNet18-based model, aiming to improve its accuracy. Our research revealed that the placement of the SCConv module significantly influences model accuracy. Consequently, we performed a series of ablation studies to pinpoint the optimal module placement, culminating in the formulation of the DdbiNet architecture. Concurrently, DdbiNet was benchmarked against several contemporary classification models, demonstrating superior accuracy in identifying hazardous driving behaviors. The ablation study results indicated that DdbiNet enhanced the Top1 and Top5 accuracy metrics by 29% and 31%, respectively, compared to the baseline model. The comparative analyses further affirmed that DdbiNet outperforms in the Top1 and Top5 accuracy metrics, while maintaining a lower count of parameters and FLOPs.