Driving fingerprinting enhances drowsy driving detection: Tailoring to individual driver characteristics

Abstract

Background

Drowsiness detection is a long-standing concern in preventing drowsiness-related accidents. Inter-individual differences seriously affect drowsiness detection accuracy. However, most existing studies neglected inter-individual differences in measurements’ calculation parameters and drowsiness thresholds. Studies without considering inter-individual differences generally used selfsame measurements and drowsiness thresholds for each participant rather than individual optimal measurements and personalized thresholds, which reduces the contribution of measurements and drowsiness detection accuracy at the individual level. Additionally, Driving Fingerprinting (DF) that represents individual traits has not been well applied in drowsiness detection.

Methods

We built the Individualized Drowsy driving Detection Model (IDDM) utilizing DF, extracting individual driver’s optimal drowsiness characteristics to detect drowsiness. Firstly, we conducted simulated driving experiments with 24 participants (2:1 male-to-female ratio, diverse ages and occupations including professional taxi drivers and graduate students) and collected data on their driving behavior, facial expressions, and the Karolinska Sleepiness Scale (KSS). Secondly, we employed a Two-layer Sliding Time Window (TSTW) to calculate DF measurements. Thirdly, we utilized attribution directed graphs to visualize DF, understand changes in DF with drowsiness, and analyze accident risks. Finally, we used DF matrices to build the IDDM. The IDDM utilized an improved adaptive genetic algorithm to extract the optimal drowsiness characteristics of individual drivers. These DF matrices, constituted by the optimal drowsiness characteristics of individual drivers, were used to train the IDDM based on principal component analysis and radial basis function neural networks. The TSTW strengthened the variation of DF with drowsiness, and the trained IDDM excavated the relationships between DF characteristics and drowsiness, which improved the accuracy and end-to-end timeliness of practical applications. The DF visualization displayed DF variations with drowsiness, theoretically supporting the use of DF to enhance personalized drowsiness driving detection.

Results

The DF visualization indicated drowsiness caused the distribution and transition probabilities of DF measurements to shift toward unsafe directions, thereby increasing the accident risk and demonstrating the rationality for utilizing DF to recognize drowsiness. The proposed IDDM achieved average accuracy, sensitivity, and specificity of 95.58 %, 96.50 %, and 94.70 %, respectively, outperforming most existing models. The trained IDDM demonstrated an average execution time of 0.0078 s and lower computational costs due to the reduction of PCA and simple RBFNN compared with models based on deep learning, and no requiring physiological data, which reduced invasiveness and enhanced feasibility for real-world implementation. The IDDM still committed challenges like integration with existing systems and concerns about privacy, which should be adjusted in implementations. This study supports the anti-drowsiness warning systems for preventing drowsiness-related accidents and promotes the integration of DF into dangerous driving research.