A domain generalization method for deploying driver distraction detection models to practical application scenarios

Driver distraction detection is crucial for reducing traffic accidents and enhancing driving safety. However, there is often a domain shift between data from real-world driving scenarios and the model training data, which limits the deployment of existing driver distraction detection models. To facilitate deployment of these models in practical application scenarios, a domain generalization method based on the contrastive language-image pretraining (CLIP) model and the constraint of center loss (DGCCL) is proposed in this paper. Firstly, the image encoder of the pre-trained CLIP model is adopted as the feature extraction module to improve the domain generalization ability of the proposed model. Furthermore, the constraint of center loss is introduced to promote the samples of different datasets to follow an approximately identical distribution in the feature space, thereby alleviating the domain shift problem. Additionally, the classification loss with additive angular margin penalty (AAMP) is introduced to further improve the cross-domain performance of the proposed model. In order to demonstrate the effectiveness of the proposed method, extensive experiments have been conducted on three publicly available driver distraction detection datasets: AUC-DDD, State-Farm, and SAM-DD. The experimental results verify that our method can achieve much better performance than various well-known models in the cross-domain driver distraction detection tasks.