Making a few talk for the many – Modeling driver behavior using synthetic populations generated from experimental data. (November 2021)

• Record Type: Journal Article
• Authors: Schindler, Ron; Flannagan, Carol; Bálint, András; Bianchi Piccinini, Giulio
• Is Part Of: Accident analysis and prevention. Volume 162(2021)
• Publication Date: 2021-11
• Subjects: Synthetic population -- Experimental data -- Bayesian functional data analysis -- Vulnerable road user -- Active safety system design; Accidents -- Prevention -- Periodicals; Accident Prevention -- Periodicals; Accidents -- Prévention -- Périodiques; 363.106
• DOI: 10.1016/j.aap.2021.106331
• Languages: English
• ISSNs: 0001-4575

Abstract

Highlights: Data from trucks in right turn maneuvers with the presence of cyclists. Modeling the behavior of a larger population of truck drivers through Bayesian methods. Enhancing the use of data sets too small in size for robust analysis. Sensitivity analysis shows impact of prior distributions. Results have practical applications for the design of future active safety systems and automated driving. Abstract: Understanding driver behavior is the basis for the development of many advanced driver assistance systems, and experimental studies are indispensable tools for constructing appropriate driver models. However, the high cost associated with testing is a serious obstacle in collecting large amounts of experimental data. This paper presents a methodology that can improve the reliability of results from experimental studies with a limited number of participants by creating a virtual population. Specifically, a methodology based on Bayesian inference has been developed, that generates synthetic cases that adhere to various real-world constraints and represent possible variations of the observed experimental data. The application of the framework is illustrated using data collected during a test-track experiment where truck drivers performed a right turn maneuver, with and without a cyclist crossing the intersection. The results show that, based on the speed profiles of the dataset and physical constraints, the methodology can produce synthetic speed profiles during braking that mimic the original curves but extend to other realistic braking patterns that were not directly observed. The models obtained from the proposed methodology have applications for the design of active safety systems and automated driving, demonstrating thereby that the developed framework has great promise for the automotive industry.