Spatio-temporal traffic anomaly detection for urban networks

Urban road networks are often affected by disruptions such as accidents and roadworks, giving rise to congestion and delays, which can, in turn, create a wide range of negative impacts to the economy, environment, safety and security. Accurate detection of the onset of traffic anomalies, specifically Recurrent Congestion (RC) and Nonrecurrent Congestion (NRC) in the traffic networks, is an important ITS function to facilitate proactive intervention measures to reduce the level of severity of congestion. A substantial body of literature is dedicated to models with varying levels of complexity that attempt to identify such anomalies. Given the complexity of the problem, however, very less effort is dedicated to the development of methods that attempt to detect traffic anomalies using spatio-temporal features. Driven both by the recent advances in deep learning techniques and the development of Traffic Incident Management Systems (TIMS), the aim of this research is to develop novel traffic anomaly detection models that can incorporate both spatial and temporal traffic information to detect traffic anomalies at a network level. This thesis first reviews the state of the art in traffic anomaly detection techniques, including the existing methods and emerging machine learning and deep learning methods, before identifying the gaps in the current understanding of traffic anomaly and its detection. One of the problems in terms of adapting the deep learning models to traffic anomaly detection is the translation of time series traffic data from multiple locations to the format necessary for the deep learning model to learn the spatial and temporal features effectively. To address this challenging problem and build a systematic traffic anomaly detection method at a network level, this thesis proposes a methodological framework consisting of (a) the translation layer (which is designed to translate the time series traffic data from multiple locations over the road network into a desired format with spatial and temporal features), (b) detection methods and (c) localisation. This methodological framework is subsequently tested for early RC detection and NRC detection. Three translation layers including connectivity matrix, geographical grid translation and spatial temporal translation are presented and evaluated for both RC and NRC detection. The early RC detection approach is a deep learning based method that combines Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). The NRC detection, on the other hand, involves only the application of the CNN. The performance of the proposed approach is compared against other conventional congestion detection methods, using a comprehensive evaluation framework that includes metrics such as detection rates and false positive rates, and the sensitivity analysis of time windows as well as prediction horizons. The conventional congestion detection methods used for the comparison include Multilayer Perceptron, Random Forest and Gradient Boost Classifier, all of which are commonly used in the literature. Real-world traffic data from the City of Bath are used for the comparative analysis of RC, while traffic data in conjunction with incident data extracted from Central London are used for NRC detection. The results show that while the connectivity matrix may be capable of extracting features of a small network, the increased sparsity in the matrix in a large network reduces its effectiveness in feature learning compared to geographical grid translation. The results also indicate that the proposed deep learning method demonstrates superior detection accuracy compared to alternative methods and that it can detect recurrent congestion as early as one hour ahead with acceptable accuracy. The proposed method is capable of being implemented within a real-world ITS system making use of traffic sensor data, thereby providing a practically useful tool for road network managers to manage traffic proactively. In addition, the results demonstrate that a deep learning-based approach may improve the accuracy of incident detection and locate traffic anomalies precisely, especially in a large urban network. Finally, the framework is further tested for robustness in terms of network topology, sensor faults and missing data. The robustness analysis demonstrates that the proposed traffic anomaly detection approaches are transferable to different sizes of road networks, and that they are robust in the presence of sensor faults and missing data.Open Acces