Identifying Unsafe Driving in Connected Vehicle Environments and its Implications for Infrastructure Providers

The goal of this research is to establish how data made available in a connected vehicle (CV) environments can benefit infrastructure providers in performing safety analysis. Specifically, the interest is to identify safety hot spots, or locations that are experiencing unexpectedly high numbers of crashes, traditionally found through methods using police-reported crash data. This process is used to evaluate transportation network safety and plan for safety-related improvements. It was expected that CV technology would provide two major improvements over the current methods, outlined by the Highway Safety Manual (HSM):

• CV has the potential to detect near-crashes (or conflicts);
• As a result, CV also has the potential to identify hot spots more proactively than current crash prediction models that rely on crash reporting, due to the availability of a much larger set of samples

The first step was to evaluate the feasibility of using existing CV datasets to identify hot spots. It was found that for this to process be successful in a connected vehicle environment, the standards must provide vehicles with a mechanism to alert infrastructure of an event that occurs and that vehicles themselves need to be able to recognize crash and near-crash situations using their on-board equipment. The focus then shifted to identifying safety-critical events, defined as crashes and near-crashes in this context, using data native to the CV standard Basic Safety Message (BSM). Three algorithms, trained using naturalistic driving study data, were proposed in three separate papers. The first was a pattern matching approach that calculated Euclidean distance between observed vehicle acceleration time series and those of some known, pre-defined actions. The algorithm saw success on a limited data set. Similarly using the same dataset, a speed prediction model was used to identify discrepancies between expected speeds and observed speeds, flagging groups of observations that were too far off from the expected speed. The third and final algorithm was trained on a much larger dataset, utilizing a discrete fourier transform and a k-means clustering algorithm to group events into clusters. This was successful on a robust dataset.

This compendium of work first, provides a comprehensive discussion of the findings related to how connected vehicle technology can benefit highway safety analyses. These findings provide a vision and a foundation for future methods and ideas as CV technology and implementation matures. Second it explores how crashes and near-crashes can be detected in connected vehicle environments. All of the hypothesized benefits of using connected vehicles for hot spot identification hinge on the ability to successfully detect crash and crash-surrogate events. As a result, a major focus of this research was modeling crashes and near-crashes in order to describe them in terms of connected vehicle data elements. Three creative methods were proposed as possible approaches to identifying these types of events, including a pattern matching approach, a speed prediction time series based approach, and a discrete fourier transform approach. Each of them have benefits and drawbacks in terms of both complexity and accuracy, but serve as excellent starting points for further research and the lessons learned are applicable and should be considered as additional models are proposed.