01628047

Component

Early detection of incidents is one of the key step to reduce incident-related congestion. With the increasing usage of GPS based navigation, promising data-scalable crowdsourced probe data is now available which can provide near-real time traffic speed information. This study utilizes such extensive historical datasets (approximately 500 GB) to gain useful insights on the normal traffic pattern of each segment. The insights come in the form of speed threshold for different time of the day and days of week for each segment. Thereafter, the anomalous traffic behaviour are classified as incidents. The dynamic thresholds developed for each segment simplifies the calibration steps that is often required when applying a model to a different dataset. Also, in this study, two alternatives of the traditional Standard Normal Deviate (SND) based incident detection algorithm are tested. The proposed algorithms can handle the masking effect of SND method where the outliers inflate the mean and standard deviation values and result in lower threshold values and in turn, lower detection rate. The high detection rate (94-97%) obtained by these algorithms compared to the SND method (83%) shows the efficacy of the models. Although higher false alarm rate (FAR) are observed for these models, but their values (4 false alarms/day) are quite lower than the acceptable FAR (10 false alarms/day) reported in previous literature.

This paper was sponsored by TRB committee ABJ30 Standing Committee on Urban Transportation Data and Information Systems.

01618707

17-05869

English

14p

2017

Transportation Research Board 96th Annual Meeting

Digital/other

Figures; References; Tables

Algorithms; Data analysis; Data files; Data mining; Incident detection

Data and Information Technology; Highways

Transportation Research Board Annual Meeting 2017 Paper #17-05869

TRIS, TRB, ATRI

Dec 8 2016 12:21PM