Driver distraction represents a road safety issue worldwide. Therefore, a distraction detection system can help improve road safety by warning drivers of non-autonomous vehicles or taking back control in the case of autonomous vehicles. Previous studies primarily relied upon classification-based supervised methods to predict distraction, but these methods require time-consuming annotation and are prone to errors. However, in this work, we propose a novel approach that explores the use of unsupervised anomaly detection techniques including traditional and deep learning methods and an anomaly score as an indicator of the distraction level. Our approach relies only on normal driving data and does not require dangerous driving situations to be collected. Therefore, this approach facilitates the acquisition of large-scale databases under real-world driving conditions. To evaluate its reliability, we use a fully annotated database containing eye tracker's data from drivers in both distracting and non-distracting settings in manual and autonomous driving modes, in a car simulator, allowing for a comparison between supervised and unsupervised models. Our unsupervised approach achieves F1-scores of 0.83 and 0.88 on the autonomous and manual driving datasets, respectively, showing close performance compared to the supervised methods. The positive correlation between anomaly scores and distraction level further demonstrates the effectiveness of our unsupervised approach.
Driver Visual Distraction Detection Using Unsupervised Learning Techniques
2023-09-24
725538 byte
Conference paper
Electronic Resource
English
METHOD FOR FINE-GRAINED DETECTION OF DRIVER DISTRACTION BASED ON UNSUPERVISED LEARNING
European Patent Office | 2023
|