Modeling Driver Car-following Behavior

Sponsored by the National Highway Traffic Safety Administration (NHTSA), the 100-car Naturalistic Driving Study was conducted in the Northern Virginia area with the recording of nearly 43,000 hours of driving data. Initially collected to investigate crash and near-crash events, this study included instrumentation of vehicles to collect and store onboard vehicle diagnostics data, global positioning systems (GPS) location information, front and rear radar tracking of objects, and synchronized video feeds viewing both the inside and the outside of the vehicle. Additional information collected from drivers includes demographics and personality questionnaires. The 100-car Study serves as the foundation for the currently ongoing SHRP 2 data collection effort, which promises to produce a much larger dataset with higher fidelity information. The current research effort focuses on the driver-specific data available from naturalistic driving studies, leveraging the unique perspective this data provides for the calibration of car-following models. Traditionally, car-following models have been both created and calibrated through the use of either loop detector data, or vehicle trajectory data created from aerial photography and videography. The data collected from these sources has limitations both in the lack of information available about the drivers, and in the length of the car-following events; limited to either instantaneous in the case of loop detector data, or as long as it takes a vehicle to progress across the field of view for the aerial trajectory data. Car-following models continue to become more sophisticated as traffic simulation software programs seek to produce more representative results compared to real-life driving behavior. It is important to probe both the limitations of the existing car-following models, and the limitations of the conventional data gathering techniques. Due to the probe-vehicle nature of naturalistic driving data, the vast dataset must first be reviewed to identify specific homogeneous sections of roadway frequently traveled by multiple drivers. Once a roadway section is identified, and a subselection of datapoints are withdrawn from the database, specific car-following events must then be identified. With respect to the 100-car study conducted not for the purpose of calibrating car-following information, but to generate safety information, this proves difficult due to the lack of sophistication in the data collection equipment, including low-resolution video feeds and sometimes unreliable object detection radar equipment. Once the processed dataset is in hand, a wealth of information relating to car-following can be investigated. The initial research investigation is intended to look at driver heterogeneity, both the variability of calibrated parameters between trips for a given driver, and the variability between drivers. Future research for this dataset may include the correlation of personality traits to model parameters, and the impact of the duration of a car-following event on the calibrated model parameters, with comparison to NGSIM data.