Learning Traffic as Images for Incident Detection Using Convolutional Neural Networks

Xiaozhou Liu,Renxin Zhong,Junzhou Chen,Hengxing Cai,Weili Sun

doi:10.1109/access.2020.2964644

Xiaozhou Liu, Renxin Zhong + Show 3 more

Open Access

https://doi.org/10.1109/access.2020.2964644

Copy DOI

Abstract

The timely and accurate detection of traffic incidents is beneficial to reduce associated economic losses and avoid secondary crashes. Inspired by the impressive success of the image classification algorithms, especially convolutional neural networks (CNNs), this study proposes a novel framework to detect highway traffic incidents by learning the traffic state as images. In such a framework, the probe vehicles equipped with the global positioning system equipment are used to obtain data. The Gramian Angular Difference Fields and Piecewise Aggregation Approximation algorithms are used to convert the link speed time series data into images. CNNs can extract the traffic features based on these images and consider an incident detection problem as a binary classification task. Further, the effectiveness of the proposed framework is evaluated by applying it to detect the traffic in a real-world environment, i.e., the Guangzhou Airport Expressway. The results illustrate that the proposed model outperforms several other algorithms with respect to almost all the performance indexes, including the detection rate with different false alarm rates and the area under the receiver operating characteristic curve.

Highlights

Traffic incidents refer to non-recurrent events that result in traffic congestion [1]
We propose a novel convolutional neural networks (CNNs) framework to transform traffic incident detection into a binary classification problem and improve the performance of certain popular methods
DATA PROCESSING This study focuses on the global positioning system (GPS) data obtained from probe vehicles, including speed and time

Summary

INTRODUCTION

Traffic incidents refer to non-recurrent events that result in traffic congestion [1]. The traffic incident detection algorithms have different data requirements, preconditions, and complexity, which are mainly divided into two categories, i.e., direct and indirect detection methods. The former method uses the video imaging technology and manual alarm reception. The proposed method can provide basic data support and decision-making basis for achieving road traffic management This necessitates a large amount of capital investment, and the current usage of surveillance videos is limited to the less-efficient manual observation. We propose a novel CNN framework to transform traffic incident detection into a binary classification problem and improve the performance of certain popular methods. The experiments illustrate that the proposed model displayed better results when compared with those exhibited by others with respect to almost all the performance indexes, including detection rate at different false alarm rates, the false alarm rate at different of detection rates, and the area under the receiver operating characteristic curve (AUC)

RELATED WORK

CONVERSION OF THE SPEED TIME

PERFORMANCE INDEXES

Findings

CONCLUSION