Development of rural curve driving models using lateral placement and prediction of lane departures using the SHRP 2 naturalistic driving data
Is Version Of
Roadway departure crashes are a major cause of fatalities on rural horizontal curves. In 2008, the Federal Highway Administration estimated that 27% of all fatalities occurred on rural highways and that among those 76% were single vehicles leaving the roadway and striking a fixed object or overturning while another 11% were head-on collisions (AASHTO 2008). Addressing crashes on rural two lane curves, specifically run off the road crashes, remains a priority for our local, state and national roadway agencies.
Much research has been conducted to look at what factors affect curve negotiation, and which factors are more likely to contribute to roadway departures. Previous research has studied how roadway factors, such as radius and shoulder width and environmental factors, such as weather affect crashes, yet limited research has been conducted looking at how driver behaviors affect crash risk. Additional research has been conducted on developing curve negotiation trajectories using small sets of curves and without much driver information.
The recent completion of the Strategic Highway Research Program 2 (SHRP 2) Naturalistic Driving Study (NDS) and Roadway Information Database (RID) allows one to expand on gaps in current literature by utilizing data from a wide variety of participants in multiple states across a broad age ranges. It also allows one to include driver factors such as age and gender, as well as drivers glance behavior and presence of distractions.
This dissertation utilizes early data from the SHRP 2 NDS and RID to develop models which provide an additional understanding of rural curve negotiation. Through three papers, two curve driving models were developed as well a model which predicts the likelihood of lane departures based off kinematic vehicle data.
In the first paper (Chapter 2) a model of normal curve driving trajectories on isolated rural two lane curves was developed using generalized least squares with an autocorrelation structure. This model found that a drivers offset 100 meters upstream of the start of the curve could help predict a vehicles position at various points throughout the curve. Additionally, the model was able to predict the average path a driver would take through seven points in the curve. These estimators suggest that drivers tend to cut the curve and are more susceptible to a lane departure at certain points in the curve.
Chapter 3, the second paper, builds on the model developed in Chapter 2 and includes additional non-isolated curves as well as non-normal driving (i.e. lane encroachments). This linear mixed effects model of curve driving trajectories included random effects for the repeated samples of drivers and drivers within the same curve as well as the same autocorrelation structure. This model was able to determine a difference in the offset at each point in the curve for those traces where a lane departure towards the inside of curve occurred and when it did not. This allowed for a boundary between normal and non-normal driving to be established. A similar correlation between the driver’s lane position upstream of the curve and lane position in the curve was also found. Smaller radii, looking down and being distracted were all found to affect trajectories in rural curves.
The final paper, Chapter 4, includes a mixed logistic regression which included a random effect for curve which took into account the repeated samples for the curves. This model produced odds-ratios for the three variables and found that increasing the amount over the advisory speed by 1 mph at the Point of Curvature (PC) of the curve increased odds of a lane encroachment towards the inside of the curve by 1.11. Shifting lane position by 0.1 m towards the inside of the curve at the PC increased odds of an inside lane departure by 1.5. In addition to the logistic regression model, two linear mixed effects models were developed which allow one to predict the speed and offset at the PC using data from 100 m upstream. This allows one to predict the probability of a lane departure 100 m upstream of the curve in addition to at the PC.