iTSS Lab at the 2018 TRB Annual Meeting

This paper presents the results of a new deep learning model for traffic signal control. In this model, a novel state space approach is proposed to capture the main attributes of the control environment and the underlying temporal traffic movement patterns, including time of day, day of the week, signal status, and queue lengths. The performance of the model was examined over nine weeks of simulated data on a single intersection and compared to a semi-actuated and fixed time traffic controller. The simulation analysis shows an average delay reductions of 32% when compared to actuated control and 37% when compared to fixed time control. The results highlight the potential for deep reinforcement learning as a signal control optimization method.

This paper investigates the application of the latest machine learning technique – deep neural networks for classifying road surface conditions (RSC) based on images from smartphones. Traditional machine learning techniques such as support vector machine (SVM) and random forests (RF) have been attempted in literature; however, their classification performance has been less than desirable due to challenges associated with image noises caused by sunlight glare and residual salts. A deep learning model based on convolutional neural network (CNN) is proposed and evaluated for its potential to address these challenges for improved classification accuracy. In the proposed approach we introduce the idea of applying an existing CNN model that has been pre-trained using millions of images with proven high recognition accuracy. The model is extended with two additional fully-connected layers of neurons for learning the specific features of the RSC images. The whole model is then trained with a low learning rate for fine-tuning by using a small set of RSC images. Results show that the proposed model has the highest classification performance in comparison to the traditional machine learning techniques. The testing accuracy with different training dataset sizes is also analyzed, showing the potential of achieving much higher accuracy with a larger training dataset.

This paper presents the results of a case study on the causes and effects of typical service disruptions in a high-speed rail (HSR) system in China – Wuhan-Guangzhou High-speed railway (WH-GZ HSR) – a 1096-kilometre HSR line. Ten months of train operation records were used to evaluate the major events causing train service disruptions or primary delay and their cascading or knock-on effects on the operations of other trains. Seven major types of delay events are identified and their impacts in terms of primary delay and number of delayed trains are analyzed. The analysis shows that, regardless the causes of the disruptions, the primary delays follow approximately similar distribution patterns. The overall impact of the disruptions, as measured by the number of trains being delayed, is shown to be largely dependent on disruption type and location. The analysis results from this research provide insights into one of the critical concerns of HSR operations – service disruptions, which is essential for developing robust train schedules and service management strategies.

This paper describes results from the project, namely, snow and ice control for parking lots and sidewalks (SICOPS) for the winter season 2016/17. The objective of this research was to field validate salt application models for parking lots developed during our previous research under semi-controlled conditions and the optimal rates proposed therein. Data was collected from 24 parking lot sites in two different cities in the U.S. – Milton, Massachusetts and Buffalo, New York. Application of the developed model to the data shows that the developed model is capable of predicting bare pavement recovery time (BPRT) with high precision for under a wide variety of weather conditions. Model performance is compromised in the presence of snow packed conditions or under conditions when the BPRT can’t be achieved with 24 hours. Deviation of model from BPRT value of zero needs further investigation.

Train operations are subject to stochastic variations, reducing service punctuality and thus quality of service (QoS). Models of such variations are needed to evaluate and predict their potential impacts for improved service management and timetabling. In this paper through a case study of the Wuhan-Guangzhou (WH-GZ) high-speed rail (HSR), we show how a wealth of train operation records can be used to model the stochastic nature of train operations at a section- and station-level. Specifically, we examine different distribution models for running times of individual sections and show that the Log-logistic probability density function is the best distributional form to approximate the empirical distribution of running times. Next, we show that the distribution of running times in each section can be used to accurately infer arrival delays. Consequently, we construct the underlying analytical model and derive the respective arrival delay distribution at the downstream stations. The results support the correctness of the presented model and that the proposed model is suitable for constructing the distribution of arrival delays at every station of the specified line. We show that the integrated distribution models of the running times and arrival delays, driven by empirical data, can be used to evaluate the quality of service (QoS) at individual track sections.

Crash prediction is a critical component of road safety analyses. A widely adopted approach to crash prediction is application of regression based techniques. The underlying calibration process is often time-consuming, requiring significant domain knowledge and expertise and cannot be easily automated. This paper introduces a new machine learning (ML) based approach as an alternative to the traditional techniques. The proposed ML model is called regularized deep belief network (DBN), which is a deep neural network with two training steps: it is first trained using an unsupervised learning algorithm and then fine-tuned by initializing a Bayesian neural network with the trained weights from the first step. The resulting model is expected to have improved prediction power and reduced need for the time-consuming human intervention. In this paper, we attempt to demonstrate the potential of this new model for crash prediction through two case studies including a collision data set from 800 km stretch of Highway 401 and other highways in Ontario, Canada. Our intention is to show the performance of this ML approach in comparison to various traditional models including negative binomial (NB) model, kernel regression (KR), and Bayesian neural network (Bayesian NN). We also attempt to address other related issues such as effect of training data size and training parameters.