New modelling approaches to analyse unsafe traffic conditions in real-time
File(s)
Author(s)
Cai, Bowen
Type
Thesis
Abstract
Unsafe traffic conditions are associated with traffic crashes and traffic conflicts. Predicting unsafe traffic conditions is a fundamental component of a proactive traffic safety management system, enabling real-time monitoring of traffic conditions, identifying complex traffic dynamics, implementing pertinent interventions and formulating policy for traffic management. This thesis utilises video analytics through deep learning models to extract dynamic traffic variables from roadside surveillance cameras and drones over freeway segments and urban junctions. Both macroscopic traffic flow variables and microscopic traffic conditions are employed to build models for predicting near future crashes and traffic conflicts.
Short-term crash prediction models are complex to develop. This is because such models should handle both excessive zeros resulting from crash counts associated with highly disaggregated observational units and temporal auto-correlation inherent in time-series crash data. To address this challenge, this PhD thesis developed a new crash prediction model termed as Zero-Inflated integer-valued Logarithmic link Time-series (ZILT) model to predict hourly traffic crashes.
Predicting traffic crashes shorter than one hour period in advance is more difficult. Classical statistical models fail to predict half-hourly crashes. Therefore, a joint model consisting of the Time-series Generalised Regression Neural Network and the Weighted CNN (WCNN) model was built. This joint model obtains the predicted values of the covariates at the next time epoch through TRGNN and makes them as the input to the WCNN model to forecast the probability of a crash event in the same time dimension.
Proactive and predictive traffic safety management system can be established based on traffic conflict and crash prediction results to monitor real-time traffic conditions. The predicted conflicts and crashes, along with the identification of influential crash affecting variables, provide instant pre-crash information to assist road control centre to conduct timely pertinent interventions to reduce hazardous conditions, prevent traffic crashes, and eventually promote roadway safety.
Short-term crash prediction models are complex to develop. This is because such models should handle both excessive zeros resulting from crash counts associated with highly disaggregated observational units and temporal auto-correlation inherent in time-series crash data. To address this challenge, this PhD thesis developed a new crash prediction model termed as Zero-Inflated integer-valued Logarithmic link Time-series (ZILT) model to predict hourly traffic crashes.
Predicting traffic crashes shorter than one hour period in advance is more difficult. Classical statistical models fail to predict half-hourly crashes. Therefore, a joint model consisting of the Time-series Generalised Regression Neural Network and the Weighted CNN (WCNN) model was built. This joint model obtains the predicted values of the covariates at the next time epoch through TRGNN and makes them as the input to the WCNN model to forecast the probability of a crash event in the same time dimension.
Proactive and predictive traffic safety management system can be established based on traffic conflict and crash prediction results to monitor real-time traffic conditions. The predicted conflicts and crashes, along with the identification of influential crash affecting variables, provide instant pre-crash information to assist road control centre to conduct timely pertinent interventions to reduce hazardous conditions, prevent traffic crashes, and eventually promote roadway safety.
Version
Open Access
Date Issued
2024-11-07
Date Awarded
01/05/2025
License URL
Advisor
Quddus, Mohammed
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)