A Fast, Scalable, Unsupervised Approach to Real-time Traffic Incident Detection

Abstract

Traffic congestion is occasionally caused by an unusual traffic incident such as a road accident or a big sporting event. The congestion could have been avoided if the traffic authority had detected and responded to it quickly and appropriately. This article explores a machine learning approach for detecting anomalous traffic incidents in real-time using GPS data collected from thousands of taxicabs in Bangkok Metropolitan area. The detection model is based on applying Principal Component Analysis (PCA) on various features extracted from overlapping fixed-length time windows over a target region. After the model has been trained, it is validated on past data and is able to discover meaningful anomalous incidents that have been verified by cross-checking with other information sources. Our approach does not require any street layout information, is computationally efficient, and can be deployed to monitor realtime traffic over large areas at scales.