Dynamic Hotspot Identification for Limited Access Facilities using Temporal Traffic Data

Publication Year:
2018
Usage 54
Downloads 32
Abstract Views 22
Repository URL:
https://stars.library.ucf.edu/etd/5777
Author(s):
Al Amili, Samer
thesis / dissertation description
Crash frequency analysis is the most critical tool to investigate traffic safety problems. Therefore, an accurate crash analysis must be conducted. Since traffic continually fluctuates over time and this effects potential of crash occurrence, shorter time periods and less aggregated traffic factors (shorter intervals than AADT) need to be used. In this dissertation, several methodologies have been conducted to elevate the accuracy of crash prediction. The performance of using less aggregated traffic data in modeling crash frequency was explored for weekdays and weekends. Four-time periods for weekdays and two time periods for weekends, with four intervals (5, 15, 30, and 60 minutes). The comparison between AADT based models and short-term period models showed that short-term period models perform better. As a shorter traffic interval than AADT considered, two difficulties began. Firstly, the number of zero observations increased. Secondly, the repetition of the same roadway characteristics arose. To reduce the number of zero observations, only segments with one or more crashes were used in the modeling process. To eliminate the effect of the repetition in the data, random effect was applied. The results recommend adopting segments with only one or more crashes, as they give a more valid prediction and less error. Zero-inflated negative binomial (ZINB) and hurdle negative binomial (HNB) models were examined in addition to the negative binomial for both weekdays and weekends. Different implementations of random effects were applied. Using the random effect either on the count part, on the zero part, or a pair of uncorrelated (or correlated) random effects for both parts of the model. Additionally, the adaptive Gaussian Quadrature, with five quadrature points, was used to increase accuracy. The results reveal that the model which considered the random effect in both parts performed better than other models, and ZINB performed better than HNB.