Data-Driven Safety Approaches for Transportation Operations and Planning

Abstract

In recent years, traffic crashes and the casualties, damages, and injuries they cause have been an increasing concern among officials and safety planners. The global incidence of motor vehicle crashes has led researchers all around the world to focus on methods both at microscopic and macroscopic levels. This dissertation mainly focuses on two levels of safety analysis: micro-level and macro-level. The micro-level is to analyze non-aggregated crash data related to traffic operations, and the macro-level safety analysis is to analyze aggregated crash data related to transportation planning. One important issue before conducting data-driven safety analysis is data quality. Raw data, including traffic safety, traffic operations, and transportation planning data require further processing to produce useful results. Among all these data sources, missing or invalid data in traffic operations data is becoming a critical concern. This dissertation proposes a systematic approach to identify and characterize traffic data error patterns to facilitate large-scale loop detector troubleshooting. Data was collected from loop detectors in Phoenix, Arizona. A set of quality control criteria was applied on daily 20-second data to find the error percentage for each loop detector. A Fuzzy C-means clustering method was implemented on the data quality check results. Then, an association rule mining method was applied to data subsets found by the clustering method to discover the most frequent rules. The analysis identified four key patterns, indicating that the proposed approach successfully found the relationships in the data errors. When the traffic data is cleaned, it could be used for a wide range of traffic safety applications. To incorporate safety analysis into traffic operations, real-time traffic data was utilized to develop crash risk prediction models by comparing the traffic statuses before each crash and several corresponding non-crash cases. In the datasets of crash prediction models, non-crash cases usually outnumber crash cases. This imbalance could affect the prediction performance of classification models. This component of the dissertation aims to examine how employing data mining techniques that account for imbalanced data could improve the predictive capability of real-time crash prediction models. Three models, including Logistic regression as the baseline, Random Forest (RF) with random under-sampling, and Adaptive Boosting (AdaBoost) were estimated with each dataset. The results were compared with the models that were estimated using the complete set of data. The results of this analysis demonstrated that using models to deal with the imbalanced data as well as lowering the variation of imbalanced data, could substantially improve crash prediction accuracy. Studying the injury severity of crashes is another component of the micro-level safety analysis. Using static safety data in the microscopic level helps to discover the cause of crashes and identify contributing factors to injury severity of crashes. Environmental factors, including adverse weather and light conditions, have been widely recognized as contributing factors to crash severity and frequency. Considering the effects of light conditions on driver perception of adverse weather, and thus on crash risk, this component of the dissertation investigates the effects of weather and light conditions on crash severity by estimating four separate multinomial logit models for specific weather (heavy rainfall or clear) and light conditions (daytime or nighttime). Marked differences were found between these conditions in terms of the significant factors affecting crash severity. In addition to traffic operations and injury severity, safety analysis needs to be conducted at the macro-level to help decision-makers proactively develop safety measures at the transportation planning level. Therefore, this component of the dissertation incorporates data-driven approaches into transportation planning. The primary purpose is to investigate the association between aggregated crashes, different trip modes, and trip purposes in Traffic Analysis Zones (TAZs). Negative Binomial (NB) and a Geographically Weighted Poisson Regression (GWPR) were estimated for total and severe crashes separately. Among different trip-related variables, home-based trips with the purpose of work and school by modes such as private vehicle, bike, walk to local bus, and park-and-ride had the most important impacts on aggregated crash data. The micro-level and macro-level safety analysis utilized in this dissertation could help safety engineers and policymakers to develop both short-term and long-term strategies to improve roadways safety. More specifically, the data-driven safety approaches conducted in this study could help to: 1- provide a framework for data quality control before using the data in safety analysis, 2- utilize the cleaned processed data for real-time crash risk prediction to help integrate safety with traffic operations, 3- analyze crash data in the micro-level to identify contributing factors to injury severity of crashes, and 4- analyze crashes in the macro-level to identify contributing factors to aggregated crashes.