Traffic incidents refer to non-recurrent events which result in traffic congestion. Drivers who encounter traffic incidents would brake or switch to other lanes, leading to an extraordinary drop in road capacity and traffic congestion. Timely and accurate detection of traffic incidents does benefit to reduce associated economic losses and avoid secondary crashes. In addition, traffic managers can rush to the scene and take appropriate measures to rescue the injured and relieve the congestion. Inspired by the great success of image classification algorithm, to be specific, the convolutional neural networks (CNN), we propose a novel framework to detect highway traffic incidents by learning traffic state as images. In this framework, probe vehicles equipped with global positioning system (GPS) equipment are used to obtain meaningful data, including link speed and time. And then, we will convert the link speed time series data into images by Gramian Angular Difference Fields (GADF) method and in conjunction with Piecewise Aggregation Approximation (PAA). Finally, CNN can extract traffic features of these images and transform the detection problem into a binary classification task (i.e. whether traffic incident occurring or not). The effectiveness of the proposed framework is evaluated by taking the real-world environment which is Guangzhou Airport Expressway and comparing with other algorithms, consisting of Support Vector Machine, Random Forests and LightGBM.

CICTP