Graduation Year

2022

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Civil and Environmental Engineering

Major Professor

Xiaopeng Li, Ph.D.

Co-Major Professor

Zhenyu Wang, Ph.D.

Committee Member

Abla Zayed, Ph.D.

Committee Member

Ankit Shah, Ph.D.

Committee Member

William Greene, Ph.D.

Keywords

Combined neural networks, Lighting illuminance, Motorcycle crash, Recursive bivariate probit, Transportation safety

Abstract

Roadway crashes have become a major cause of human deaths and injuries and caused much economic damage. According to World Health Organization (WHO), about 1.35 million people died in traffic crashes and over 50 million people injured from traffic crashes in the year of 2016 (World Health Organization, 2020). Nationally, in the U.S., the latest decade witnessed an increase of over 10% in traffic fatalities (National Highway Traffic Safety Administration (NHTSA), 2019). Similarly, in state of Florida, the fatal crashes increased by 4.62% from 2019 to 2020 (Florida Department of Highway Safety and Motor Vehicles, 2021). To prevent injuries and reduce the economic loss due to traffic crashes, scholars and researchers has been proposing innovative approaches to mitigate traffic crashes.

In all crashes, nighttime crashes are over-represented on the US highway system. According to the National Highway Traffic Safety Administration (NHTSA) Traffic Safety Facts 2015, 51% of fatal crashes and 29.5% of injury crashes occurred at night (dark, dark but lighted, dawn/dusk); meanwhile only 21–23% of the vehicle miles traveled (VMT) were at night (Monsere and Fischer, 2008). Roadway lighting, which provides additional visibility by supplementing vehicle headlights, has been identified as an effective countermeasure to improve nighttime safety. However, the effects of street lighting photometric parameters in reducing nighttime crashes on roadway segments are not well understood in the literature. To bridge this gap, this dissertation investigated the effects of street lighting illuminance (rather than the presence of street lighting) on nighttime crash occurrence on roadway segments and developed a Crash Modification Function (CMF) for horizontal illuminance on roadway segments.

Illuminance data were collected from 440 roadway segments (each segment being the roadway between two successive signalized intersections) in Florida, from 2012–2014 using the Advanced Lighting Measurement System (ALMS). Four years of nighttime and daylight crash data (2011–2014) were matched to the selected segments. Based on the collected data, nighttime and daylight random parameter negative binomial (RPNB) models were estimated to address unobserved heterogeneity over segments. The expected night-to-day crash odds ratio (equivalent to CMF) was calculated based on the RPNB models to quantify the change in relative risk of nighttime crashes caused by an alteration of horizontal illuminance on roadway segments. The bootstrap resampling technology was used to estimate the 95% confidence interval of CMF. Mean horizontal illuminance was identified to be normally distributed with a mean of -0.126 and a standard deviation of -0.094 in the nighttime model. The randomness captures the effect of unobserved confounders related to mean horizontal illuminance—increases in mean horizontal illuminance tend to reduce nighttime crashes on 91% of segments and increase nighttime crashes on the remaining 9% of segments. An expected daylight-to-day odds ratio-based CMF was developed as a power function of mean horizontal illuminance over a baseline. Other significant variables contributing to nighttime crash risk include illuminance standard deviation, Annual Average Daily Traffic (AADT), truck percentage, segment length, access density, undivided road type, and urban/city limits. However, the correlations between roadway illuminance average and uniformities are discovered. Thus, the dissertation utilized a matched-case control method to quantify the effects of illuminance uniformity on roadway nighttime crash occurrences. Illuminance data were collected from 300+ center miles of roadway segments in Florida from 2012–2014 and matched four years of nighttime crashes. The measured roadway corridors were split into uniform segments with a length of 600 ft. Each uniform segment was labeled as a case (with nighttime crashes that occurred in a certain year) or a control (without nighttime crashes that occurred in a certain year). To control excess zero crash observations, confounding effects, and temporal variation, cases was matched to control at a ratio of 1:1 by illuminance mean, which is strongly correlated to both illuminance standard deviation (uniformity) and nighttime crash frequency, in each year. A conditional logistic model was fitted based on 1,785 paired case-control stratums to estimate the relative risk of nighttime crashes due to changes in street lighting uniformity. Based on the model, significant (at a confidence level of 95%) and consistent crash modification factors (CMFs) for illuminance uniformity were produced; compared to illuminance standard deviation less than 0.2, illuminance standard deviation between 0.2 and 0.57, illuminance standard deviation between 0.57 and 0.7, and illuminance standard deviation more than 0.7 experienced 1.32, 1.42, and 2.42 times the risk of nighttime crashes along roadway corridors.

Also, motorcycle crashes on horizontal curves are also lack of attention in the U.S. Texas Motorcycle Crash Facts (Texas DPS, 2018) indicate that motorcyclists experience fatal and severe crashes eight times more than non-motorcyclists on curves. A cause-effect chain, which describes the relationship between contributing factors, driver/rider improper pre-crash actions, and crash outcome (injury severity), exists in motorcycle-vehicle crashes on horizontal curves. Previous studies did not address the correlation between injury severity and improper actions in identifying risk factors. This study aimed to develop a recursive bivariate analysis to simultaneously investigate the effects of covariates on motorcyclist fatality and improper actions (for both riders and drivers) in curve-related motorcycle-vehicle crashes. Two recursive bivariate probit models were developed to identify significant factors that contribute to riders' or drivers' improper actions, factors that directly impact motorcyclist fatality only, and factors that influence motorcyclist fatality and riders' or drivers' improper actions simultaneously. The direct, indirect, and joint marginal effects of the identified contributing factors on motorcyclist fatality risk were addressed based on fitted models. It is indicated that either riders' or drivers' improper actions in a motorcycle-vehicle crash significantly increase motorcyclist fatality risk. Riders' physical defects and alcohol/drug involvement are the most significant factors contributing to both riders' improper pre-crash actions and motorcyclist fatality. Curve design features were also found to have significant but diverse impacts on rider/driver improper actions and/or motorcyclist fatality risk. Other significant factors included roadway, rider, and driver characteristics. The recursive bivariate probit analysis approach produced fruitful results and provided useful information about motorcycle crash causations.

Crash hot spot identification helps highway safety management more efficiently to predict the roadway sections with high crash risks and implement improvement measures on the precise locations with limited resources. This research uses a microscopic spot (e.g., a square grit unit) to capture the detailed crash pattern and suggest more accurate safety countermeasures. Further, for the first time, the study proposed a novel identification technique by integrating unstructured data (satellite images) and structured data (recorded roadway features and traffic information). Both data are expected to affect the safety level of a specific roadway spot. A combined neural network model is developed with a multiple perception layer neural network (MLP) extracting features from structured data and a convolutional neural network (CNN) dealing with the satellite images. The representations from the two branches (MLP and CNN) are concatenated to perform the final representation of crash risks which is used to make predictions. This proposed model successfully identifies crash hot spots with an accuracy of 88%, which significantly higher than other machine learning models using either structured data or satellite images. This proposed mixed data model can help transportation agencies and planners to predict the crash locations in advance to avoid time-consuming investigation and save costs on roadway safety diagnosis across different roadways and intersections.

This dissertation identified the two high-risk patterns in crash analysis, nighttime crash and motorcycle crashes and utilized economic models to quantify the effects of lighting on nighttime crash risks and develop the cause-and-effect chains for motorcycle crash fatalities to prevent them. Further, the dissertation proposed a combined neural network using both traditional traffic data and satellite images to identify the high-risk road spot in the roadway network.

Share

COinS