Approximating incident occurrence time with a change-point latent variable framework
Authors
Abstract
We propose a methodology to approximate actual incident occurrence time by analyzing downstream volume sensor data. We model the time difference between actual occurrence time and reported time (or \emph{delay}) as a latent variable that becomes a parameter in a change-point time series model. We then apply a \emph{maximum a posteriori} (MAP) framework to infer the most probable delay. This MAP framework uses the time series model as the likelihood function and a bayesian prior based on field knowledge.We applied our model on 5 months of traffic sensor data and accident reports from 3 Singapore expressways and corrected the accident start times for 1086 accidents in total. We compared the results with a manually constructed baseline and obtained a mean absolute error (MAE) between 5.7 and 7.4 minutes and a root mean squared error (RMSE) between
10 and 12.