Investigating the Relationship Between Latent Driving Patterns and Traffic Safety Using Smartphone-Based Mobile Sensor Data

INVESTIGATING THE RELATIONSHIP BETWEEN LATENT DRIVING PATTERNS AND TRAFFIC SAFETY USING SMARTPHONE-BASED MOBILE SENSOR DATA Kenneth Wynne Old Dominion University, 2016 Director: Dr. Rajesh Paleti Crash frequency modelling has been used in the past as an attempt to quantify the expected number of crashes occurring on a certain segment of roadway given a set of variables and factors describing the roadway segment and the traffic along that segment. These models are referred to as the Safety Performance Functions (SPFs) in the Highway Safety Manual (HSM). In past studies, these SPFs have focused primarily on roadway geometric information along with limited traffic exposure data such as traffic volume. Alternate data sources for probe vehicle data are increasingly available and this research sought to exploit this new information in order to obtain an improved model. Specifically, this research aims to make use of the accelerometer sensors in smartphones to extract microscopic traffic measures that can serve as better indicators of driving patterns. The study focused on crash frequency along roadway segments in the Hampton Roads region. To start-off, mobile sensor data was collected by driving along major roadways in the Hampton Roads region during the evening peak period (4 to 6 pm). Next, this data was overlaid on the transportation network to map probe data and the roadway segments. Then, several acceleration and deceleration metrics were calculated for each roadway using the mobile sensor data. Subsequently, these metrics were appended to the VDOT crash data for the past one year. Supplementary data sources were used to assemble information regarding roadway inventory data and traffic exposure information. Next, statistical model estimation was undertaken to identify the factors affecting crash frequency along major interstates in Hampton Roads. The results indicate that when comparing a model based solely on roadway geometrics to a model including both roadway geometrics and probe vehicle data, the combined model was a significant improvement. Several probe vehicle data parameters capturing microscopic traffic conditions were significant in the final model. Lastly, elasticity analysis was undertaken to quantify the relative impact of different factors in the model. With regard to statistical modeling, this research considered both a Poisson and a negative binomial model that served as standard models for crash frequency modeling in the literature. The negative binomial model was found to be a significant improvement over the Poisson model. Previous research has indicated that negative binomial models tend to perform better than Poisson models when there is overdispersion present in the dataset. This research supports this claim. Overall, this research has determined that the addition of probe vehicle data to roadway inventory data and the usage of a negative binomial model have proved to provide a robust crash frequency model.