Citation

Xie K, Ozbay K, Yang H, Li C. Transp. Res. C Emerg. Technol. 2019; 106: 61-72.

Copyright

(Copyright © 2019, Elsevier Publishing)

DOI

10.1016/j.trc.2019.07.004

PMID

unavailable

Abstract

Proactive safety management has gained increasing attention for its potential to mitigate crash risk. This study aims to leverage massive vehicle trajectory data for proactive safety analytics. State-of-the-art computer vision techniques are employed to automatically extract massive vehicle trajectories from 70-hour traffic video data at two intersections in Brooklyn, New York City. The novelty of our trajectory extraction algorithm includes the inclusion of high-level information from foreground/background separation to cluster feature points that belong to the same vehicle and the use of non-parametric clustering method Dirichlet process Gaussian mixture model (DPGMM) that does not require specification of the cluster number. The amount of video data used in this case study is substantially greater than what was previously used in most of literature. Surrogate safety measures in terms of time to collision are introduced to identify rear-end conflict risk for adjacent vehicles. Hidden Markov models (HMMs) are then proposed to model the rear-end conflicts at five-minute intervals. The proposed HMMs are found to have better performance in terms of representing the conflict occurrence and their predictive abilities are comparable to the classical autoregressive integrated moving average (ARIMA) models. HMMs are then used to infer the hidden states of traffic safety. As a result, frequent switches between different states and a clustering of high-risk states are observed. The modeling results imply that HMMs can help monitor the prevailing traffic conditions and facilitate proactive safety management.

Language: en

Keywords

Artificial intelligence; Hidden Markov model; Proactive safety management; Safety surrogate measures; Traffic video; Vehicle trajectories