Spatio-temporal Learning for Traffic Events

Abstract

The study on traffic events has been an important topic in urban studies. Traffic events, ranging from disabled vehicles to abnormal driving behaviors at a local scale, to extreme events such as hurricanes or large civic events at the city scale, can cause anomalous traffic conditions. Accurate and timely detection of such traffic events can facilitate services ranging from emergency response to congestion mitigation, and ultimately help transportation agencies make informed decisions. With the deployment of advanced sensing and communication technologies that are enabling massive volumes of new data to be available, there is now potential to automate anomaly detection processes via machine learning and deep learning techniques. To address the problems related to detecting events in urban traffic, this dissertation develops a series of data-driven models, ranging from high-dimensional linear models to non-linear deep learning models. The proposed models consider the spatio-temporal variations in transportation, while addressing the data sparsity issues that are common in traffic datasets. The main contributions of the thesis are as follows: i) For large event detection, a robust tensor recovery algorithm is developed that can impute moderate volumes of missing data and detect outliers at the same time. ii) The proposed robust tensor recovery method is extended to an online version to handle large streaming datasets. iii) To take full advantage of the road network structure while also exploring the non-linear traffic patterns, a graph neural network-based autoencoder is proposed for traffic event detection. iv) For local traffic event detection, a spatial-temporal model for highway anomaly detection via a recurrent graph attention autoencoder is developed. The model can detect abnormal driving behaviors that violate commonly accepted norms or do not conform to the behavior of other nearby drivers.