Machine learning based online traffic incident detection and management for urban networks

• Main Author: Yang, Huan
• Other Authors: Wang Dan Wei
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Control engineering; Engineering::Civil engineering::Transportation
• Online Access: https://hdl.handle.net/10356/153158
• Institution: Nanyang Technological University

Urban traffic networks are often choked due to non-recurrent incidents at random locations. Heavy economic costs, environmental pollution, and severe noise pollution arise from the delay of incident detection and the lack of alternative traffic management strategies. Therefore, it is of great significance to detect the incident and identify the location timely. Afterwords, deploying an efficient Traffic Signal Control (TSC) strategy for the intersections affected by the incident based on traffic information such as traffic volume and average speed is also necessary. However, because of the complexity of urban networks, many traditional automatic incident detection (AID) methods in freeway networks may become cumbersome in urban networks. Similarly, adaptive TSC strategies for normal traffic conditions are inapplicable for the control tasks with traffic incidents. In view of these, this thesis investigates online traffic incident detection and management problems for urban networks. First of all, a dynamic origin-destination (OD) scheme is developed for further studies in traffic incident detection and management. As the input of the dynamic traffic assignment (DTA) simulation model, dynamic OD matrix represents the traffic demand in a specific time period and is crucial for intelligent transportation applications. In order to develop the online incident detection and management framework, simulation models with near-to-real dynamic traffic demand need to be established for testing and validation purposes. Since traffic flow data collected by Inductive Loop Detectors (ILDs) cannot be directly used in the simulator, we must develop a dynamic OD estimation method to transform the traffic flow data into the traffic demand. For dynamic OD estimation, most existing methods use historical OD matrices as the initial state and then calibrate them to get the result. However, the acquisition of historical OD matrices is costly and it may lose its reference value as the traffic network changes. Hence, this thesis develops an effective dynamic OD estimation approach optimization that does not need to use the historical OD matrix. Through a bi-level optimization structure, OD matrices of different time intervals are estimated sequentially. A ridge regression method is applied to obtain the initial OD matrix. A constrained nonlinear programming method is proposed to calibrate the assignment matrix for accurately mapping the demand to the traffic flows. With the assignment matrix, an LSQR based algorithm and a modified simultaneous perturbation stochastic approximation (SPSA) algorithm, called Restart-SPSA, are proposed to estimate OD matrix in each time interval. The capability of the proposed solution scheme has been validated in two urban areas of Singapore using the microscopic traffic simulator VISSIM . Secondly, two feasible online AID approaches are developed. The one is an online Convolutional Neural Network (CNN) based traffic incident detection method using three types of traffic flow data (traffic volume, speed and acceleration of vehicles). Regarding the traffic flow data during a time period as an image with RGB channels, this method employs the idea of CNN to deal with image classification. The online incident detection is realized by using the time sliding window technique. Experiments conducted via a microscopic simulation platform show that the proposed CNN-based method can effectively and efficiently detect even minor incidents in urban networks and the acceleration data has the biggest impact on the detection accuracy. Compared with LSTM and BP methods, the CNN-based method also demonstrates its high performance. The other is a novel unsupervised learning based automatic incident detection (AID) method using traffic flow data collected by Inductive loop detectors (ILDs). A Long Short Term Memory - Variational Autoencoder (LSTM-VAE) model is proposed to extract nonlinear features of traffic flow data with strong spatial-temporal correlations and build the incident-free model offline. In the online monitoring part, two statistics named z^2 and SPE are constructed to detect the occurrence time of traffic incidents. Subsequently, the contribution plot technique is employed to localize the detected incidents. Through comparing with baseline methods including California Algorithm series, Principal Component Analysis (PCA), Autoencoder(AE), and Variational Autoencoder(VAE), it has been demonstrated that the efficiency, effectiveness, and transferability of the proposed method in both freeways and urban networks. Finally, a traffic incident management (TIM) approach is proposed to prevent congestions by adjusting the traffic signals of intersections near the incident point. In theory, traffic signal control can be regarded as an optimization problem, which tries to cut down the congestion level or reduce the travel delay time. By denoting the traffic flow data as the state, traffic signal schedule of a cycle as the action, and a weighted average delay as the reward, this problem can be solved through deep reinforcement learning (DRL) approaches. In this thesis, a DRL algorithm called Twin Delayed Deep Deterministic Policy Gradient (TD3) is employed to control the phase splits under normal traffic conditions. However, a well-trained DRL agent may have difficulties when facing suddenly changed environments caused by incidents. Thus, to cope with such unexpected changes in both traffic capacity and demand, a genetic algorithm based gradient-free optimizer called GA-critic is proposed to generate the optimal phase splits under such conditions. Particularly, when an incident is detected and the corresponding road segment is closed for route diversion, two online TIM strategies are proposed to prevent congestion. Several implemented experiments have showcased the high control performance of the developed TD3-based TSC strategy and the effectiveness of the online TIM strategies.