Event-related driver stress detection with smartphones among young novice drivers

Abstract

Complex and diverse driving situations can pose short-term stressors to novice drivers. Continuously detecting stress is essential for driver training, stress intervention, and the design of in-vehicle information systems. This study designed and validated a driver stress detection method at the event level based on machine learning algorithms and facial features captured with smartphones. Thirty young novice drivers completed two driving tasks containing eight events of two versions (neutral and stressful), with psychological, physiological, and facial data collected. Four combinations of input data types and six machine learning algorithms were used to detect stressful events. The KNN algorithm with facial plus individual profile features yielded the highest accuracy of 89.2%. Adding individual profile features can improve classification performance. Facial areas such as brow, eye, jaw, nose, and mouth were most sensitive to stress. This approach could provide more temporal-spatial information about the driver’s stress levels during the whole driving process. Practitioner Summary: This paper proposed a method to detect driver stress at the event level with smartphones. Models with facial plus individual profile features and the KNN algorithm had the most outstanding classification performance. The presented approach can serve as a tool for improving in-vehicle interaction system design when considering driver stress.

Abbreviations: GSR: galvanic skin response; ECG: electrocardiography; HR: heart rate; HRV: heart rate variability; RGB: red green blue; NIR: near-infrared; IP: individual profile; DSI: driver stress inventory; APS: arousal predisposition scale; API: application programming interface; PPG: photoplethysmography; EDR: electrodermal response; PD: pupil diameter; SCL: skin conductance level; RF: random forest; KNN: k-nearest neighbour; LDA: linear discriminant analysis; QDA: quadratic discriminant analysis; SVML: support vector machines with the linear kernel; SVMP: support vector machines with the polynomial kernel; TP: true positive; TN: true negative; FP: false positive; FN: false negative; t-SNE: t-distributed stochastic neighbour embedding

Keywords:

• Driver stress
• driving event
• facial expression
• classification
• smartphone

Additional information

Funding

This study was supported by the National Natural Science Foundation of China under grant number 71771132 and the National Key R&D Program of China (2018YFB1600500).