Low-intrusive Driving Anger Classification Method Based on Semi-supervised Learning
To monitor anger while driving in real time and provide timely and effective intervention and adjustment,an accurate and efficient method for classifying anger while driving is proposed.Based on low-intrusive driving behavior and voice features,this study adopted semi-supervised learning methods to build a model to reduce the dependence on labels and improve classification accuracy.The driving data were obtained from a high-fidelity driving simulation experiment involving 30 participants.A sliding time window was set to intercept anger events,and a driving anger dataset was formed through feature extraction and computation.On this basis,a model called SSL-GBM was developed by combining a pseudo-labeling algorithm in semi-supervised learning(SSL)with a gradient boosting machine(GBM),thus fully exploring the internal information of the data to reduce the dependence on manual labels.Data processing,feature engineering,model searching,and parameter optimization were automated within an automated machine framework,enabling the classification of driving anger levels.The results indicate that the driving anger emotion classification model has an accuracy of 90.3％in predicting five-level driving anger scores,which is an improvement of 3.7％compared to the best-performing model among the existing models.In particular,the recognition accuracy for levels 2-5 improves by more than 2.5％,significantly reducing the detection failure to misjudge the angry state as normal.As shown by the prediction of anger levels throughout the driving duration,the algorithm is fully equipped with the characterization ability and generalization performance applied to real-time driving anger state recognition,thereby verifying the effectiveness and rationality of the proposed approach.This study has significant application value in discriminating driving anger states and enhancing the capacity of driving assistance systems to monitor dangerous driving behaviors.
NETL
NSTL
万方数据
