The past decade witnessed rapid developments in traffic data sensing technologies in the form of roadside detector hardware, vehicle on-board units, and pedestrian wearable devices. The growing magnitude and complexity of the available traffic data has fueled the demand for data-driven models that can handle large scale inputs. In the recent past, deep-learning-powered algorithms have become the state-of-the-art for various data-driven applications. In this research, three applications of deep learning algorithms for traffic state estimation were investigated. Firstly, network-wide traffic parameters estimation was explored. An attention-based multi-encoder-decoder (Att-MED) neural network architecture was proposed and trained to predict freeway traffic speed up to 60 minutes ahead. Att-MED was designed to encode multiple traffic input sequences: short-term, daily, and weekly cyclic behavior. The proposed network produced an average prediction accuracy of 97.5%, which was superior to the compared baseline models. In addition to improving the output performance, the model's attention weights enhanced the model interpretability. This research additionally explored the utility of low-penetration connected probe-vehicle data for network-wide traffic parameters estimation and prediction on freeways. A novel sequence-to-sequence recurrent graph networks (Seq2Se2 GCN-LSTM) was designed. It was then trained to estimate and predict traffic volume and speed for a 60-minute future time horizon. The proposed methodology generated volume and speed predictions with an average accuracy of 90.5% and 96.6%, respectively, outperforming the investigated baseline models. The proposed method demonstrated robustness against perturbations caused by the probe vehicle fleet's low penetration rate. Secondly, the application of deep learning for road weather detection using roadside CCTVs were investigated. A Vision Transformer (ViT) was trained for simultaneous rain and road surface condition classification. Next, a Spatial Self-Attention (SSA) network was designed to consume the individual detection results, interpret the spatial context, and modify the collective detection output accordingly. The sequential module improved the accuracy of the stand-alone Vision Transformer as measured by the F1-score, raising the total accuracy for both tasks to 96.71% and 98.07%, respectively. Thirdly, a real-time video-based traffic incident detection algorithm was developed to enhance the utilization of the existing roadside CCTV network. The methodology automatically identified the main road regions in video scenes and investigated static vehicles around those areas. The developed algorithm was evaluated using a dataset of roadside videos. The incidents were detected with 85.71% sensitivity and 11.10% false alarm rate with an average delay of 27.53 seconds. In general, the research proposed in this dissertation maximizes the utility of pre-existing traffic infrastructure and emerging probe traffic data. It additionally demonstrated deep learning algorithms' capability of modeling complex spatiotemporal traffic data. This research illustrates that advances in the deep learning field continue to have a high applicability potential in the traffic state estimation domain.

Smart cities are revolutionizing the transportation infrastructure by the integration of technology. However, ensuring that various transportation system components are operating as expected and in a safe manner is a great challenge. One of the proposed solutions is traffic monitoring systems which collect and analyze traffic data for the safe operation and management of the overall system. Even though traffic safety analysis has been tied to crash data, surrogate safety measures (SSM) have recently emerged as a replacement. SSM can provide a convenient alternative for understanding the impact of conflicts on overall road safety. Traditionally, conflicts were studied through manual methods that are time-consuming and prone to biases. With the onset of new technologies such as automated video techniques, micro-simulation and Vehicular ad-hoc Networks researchers are adopting statistical and machine learning methods for developing reliable indicators for possible traffic crashes. However, these technologies don't provide high-level reasoning about the overall state of the traffic. In this work, we propose the use of formal methods as a means to specify and reason about the traffic network's complex properties. Formal methods provide a flexible tool to define the safe operation of the traffic network by capturing non-conforming behavior, exploring various possible states of the traffic system and detecting any inconsistencies within it. Hence, we develop specification-based monitoring for the analysis of traffic networks using the formal language, Signal Temporal Logic. We develop monitors that identify safety-related behavior such as conforming to speed limits, maintaining appropriate headway, and performing safe lane-change maneuvers. The framework is tested using a calibrated micro-simulated highway scenario and offline monitoring is applied to individual vehicle trajectories to check whether they violate or satisfy the defined safety specifications. Statistical analysis of the outputs show that our approach can differentiate violating from conforming vehicle trajectories based on the defined specifications. This work can be utilized by traffic management centers to study the traffic stream properties, identify possible hazards, and provide valuable feedback for automating the traffic monitoring systems.

The importance of sensing technologies in the field of transportation is ever increasing. Rapid improvements of cloud computing, Internet of Vehicles (IoV), and intelligent transport system (ITS) enables fast acquisition of sensor data with immediate processing. Machine learning algorithms provide a way to classify or predict outcomes in a selective and timely fashion. High accuracy and increased volatility are the main features of various learning algorithms. In this dissertation, we aim to use infrastructure- and vehicle-based sensors to improve safety and mobility of urban transportation systems. Smartphone sensors were used in the first study to estimate vehicle trajectory using lane change classification. It addresses the research gap in trajectory estimation since all previous studies focused on estimating trajectories at roadway segments only. Being a mobile application-based system, it can readily be used as on-board unit emulators in vehicles that have little or no connectivity. Secondly, smartphone sensors were also used to identify several transportation modes. While this has been studied extensively in the last decade, our method integrates a data augmentation method to overcome the class imbalance problem. Results show that using a balanced dataset improves the classification accuracy of transportation modes. Thirdly, infrastructure-based sensors like the loop detectors and video detectors were used to predict traffic signal states. This system can aid in resolving the complex signal retiming steps that is conventionally used to improve the performance of an intersection. The methodology was transferred to a different intersection where excellent results were achieved. Fourthly, magnetic vehicle detection system (MVDS) was used to generate traffic patterns in crash and non-crash events. Variational Autoencoder was used for the first time in this study as a data generation tool. The results related to sensitivity and specificity were improved by up to 8% as compared to other state-of-the-art data augmentation methods.

This thesis contributes to our understanding of the changes in traffic volumes on major roadway facilities in Florida due to COVID-19 pandemic from a spatiotemporal perspective. Three different models were tested in this study- a) Linear regression model, b) Spatial Autoregressive Model (SAR) and c) Spatial Error Model (SEM). For the model estimation, traffic volume data for the year 2019 and 2020 from 3,957 detectors were augmented with independent variables, such as- COVID-19 case information, socioeconomics, land-use and built environment characteristics, roadway characteristics, meteorological information, and spatial locations. Traffic volume data was analyzed separately for weekdays and holidays. SEM models offered good fit and intuitive parameter estimates. The significant value of spatial autocorrelation coefficients in the SEM models support our hypothesis that common unobserved factors affect traffic volumes in neighboring detectors. The model results clearly indicate a disruption in normal traffic demand due to the increased transmission rate of COVID-19. The traffic demand for recreational areas, especially on the holidays, was found to have declined after March 2020. In addition, change in daily COVID-19 cases was found to have larger impact on South Florida (District 6)'s travel demand on weekdays compared to other parts of the state. Further, the gradual increase of traffic demand due to the rapid vaccination was also demonstrated in this study. The model system will help transportation researchers and policy makers understand the changes in traffic volume during the COVID-19 period as well as it's spatiotemporal recovery.

Rapid growth in population along with urban-centric activities impose a massive demand on existing transportation systems, thus increasing traffic congestion and other mobility related challenges. To overcome such challenges, we need network-scale models to accurately predict real-time traffic demand and associated congestion. However, traditional network modeling approaches have shortcomings due to the complexity in traffic flow modeling, limited scope to incorporate real-time data available from emerging data sources and requiring excessive computation time to generate accurate estimation of traffic flows. Advancement in traffic sensing technologies with big data has created a new opportunity to overcome these challenges and implement deployable data-driven models to predict network-level traffic dynamics and congestion propagation in real time. However, existing data-driven approaches are limited in scope: they are developed for small-scale networks; they do not consider the fundamental concept of traffic flow propagation; and they are applied for short-term prediction ( < 1 hour). In this dissertation, we develop graph convolution based neural network architectures for network scale traffic modeling as a solution to overcome these limitations. First, we develop a Graph Convolutional Neural Network (GCNN) Model to solve the traffic assignment problem in a data-driven way; the validation results show that the model can learn the user equilibrium traffic flow well (mean error < 2%). Since the model can instantaneously determine the traffic flows of a large-scale network, this approach can overcome the challenges of deploying mathematical programming or simulation-based traffic assignment solutions for large-scale networks. Second, we scale this approach and develop a Graph Convolutional LSTM (GCN-LSTM) model for traffic movement volume prediction at intersection level. We rigorously tested the model over traffic movement volume data collected from Seminole County's automated signal performance measure (ATSPM) database which show that 90% of cases, absolute error of the predicted values is less than 20. Finally, we develop a Dynamic Graph Convolutional LSTM (DGCN-LSTM) model to predict evacuation traffic flow for interstate network of Florida. The implemented model can be applied to predict evacuation traffic over a longer forecasting horizon (6-hour) with higher accuracy (R^2score 0.95). Hence, it can assist transportation agencies to activate appropriate traffic management strategies to reduce delays for evacuating traffic.

This research presents an approach for safety diagnosis using sensor fusion techniques. This work fuses the outputs of a roadside low-resolution camera and a solid-state LiDAR. For vehicle classification and detection in videos, the YOLO v5 object detection model was utilized. The raw 3D point clouds generated by the LiDAR are processed by two manual steps - ground plane transformation and background segmentation, and two real-time steps - foreground clustering, and bounding box fitting. Taking the generated 2D bounding boxes of both camera and LiDAR, we associate the common bounding box pairs by thresholding on the Euclidean distance threshold of 6 ft between the centroid pairs. We perform weighted measurement update based on the RMSE of each of the sensor's detection compared to manually labeled ground truths. The fused measurements are tracked by using linear constant velocity Kalman Filter. With the generated trajectories, we compute post encroachment time (PET) at pixel level conflicts based on the generated vehicle trajectories. We have proposed a complete bipartite graph matching strategy of vehicle parts along with the conflict angle to obtain conflict types - rear-end, sideswipe, head-on, and angle conflict. A case study on a signalized intersection is presented. The output of the proposed framework performs significantly better than the single sensor-based systems in terms of the number of detections and localization. It is expected that the proposed method can be employed to diagnose road safety problems and inform the required countermeasures.

This study proposes a software to upgrade the UCF SST's Automated Roadway Conflicts Identification System (ARCIS), a pixel-to-pixel manner automated safety diagnostics and conflict identification system. The system is developed to extract vehicles' trajectories and traffic parameters using unmanned aerial vehicles (UAV) video and utilizing deep learning techniques. A user-friendly tool to improve rapid system development with active-learning, data analysis, and visualization techniques is introduced, which is capable of traffic safety near-miss diagnostics based on the ARCIS output. Multiple approaches are used to enhance the system performance, including video stabilization, object filtering, stitching multiple videos, vehicle detection and tracing. In addition, the active learning technique based on Stream-Based Selective Sampling strategy is adopted for a human-in-the loop label correction that is developed in order to reduce the labeling time and cost. The system outputs 3D maps of vehicle speed, count and surrogate safety measures, which provide insights for traffic safety diagnosis. Ultimately, these functionalities were integrated into a comprehensive system for traffic safety applications. Previous studies only investigated methods for enhancing road traffic safety and traffic network data analysis; this study builds upon the literature but improves upon it with an efficient video processing methodology, a higher quality and accuracy result on traffic trajectory data, and the ability to visualize the data in various formats for traffic analysis.

Diverging diamond interchanges (DDIs) and Displaced left-turn intersections (DLTs) are designed to enhance the operational performance of conventional intersections that are congested due to heavy left-turn traffic volumes. Since drivers are not familiar with these types of intersections, there is a need to evaluate their safety performance to validate their effect, and to estimate reliable and representative Crash Modification Factors (CMFs). The safety evaluation was conducted based on three common safety assessment methods, which are before-and-after study with comparison group, Empirical Bayes before-and-after method, and cross-sectional analysis. Furthermore, since DLTs showed poor safety performance, the study also investigated the operational performance of DLTs using a general linear model describing the relationship between traffic delay and other operational and geometric characteristics based on high-resolution traffic data. The DDI analysis included a sample size of 80 DDIs and 240 conventional diamond interchanges in 24 states, while the DLT analysis included 13 DLTs and 26 conventional intersections in 4 states. The analysis results indicated that converting conventional diamond interchanges to diverging diamond interchanges could significantly decrease the total, fatal-and-injury, rear-end and angle/left-turn crashes by 26%, 49%, 18%, and 68%, respectively. On the other hand, converting conventional intersections to displaced left-turn intersections could significantly increase the total number of crashes as well injury crashes and some other crash types (i.e., single vehicle, angle). However, the operational analysis implied that they have the potential to reduce the delay at intersections by 3.567 sec/veh. Consequently, the study quantified the costs and benefits associated with implementing DLTs. The results showed that this alternative design could provide much benefits in terms of its operational performance. However, its poor safety performance could result in losses much higher than its benefits. The study concludes that DDIs could significantly decrease crash frequency, while DLTs could not provide safety benefits. However, DLTs might be more efficient for operational performance. It is recommended that appropriate safety countermeasures should be developed and implemented to enhance traffic safety at DLTs.

The ubiquity of smartphones has made a remarkable influence on everyone's day to day life. Variety of useful built-in sensors provide smartphones with a convenient floor for data collection and analysis. Application development based on the user's location and movement is not a difficult task nowadays. But injuries and deaths due to smartphone-distracted movement on roadways is on the increase. This study explores the capabilities of smartphone inertial sensors for pedestrian activity recognition. Smartphone distracted movements can be predicted from the associated pedestrian's posture, thus inertial sensors can provide effective solution for this specific task. Volunteers were asked to perform different pedestrian activities with smartphones in their hand or in trouser pocket. Accelerometer and gyroscope data were collected, and time windowing was applied for proper segmentation of the data. After time and frequency domain feature extraction of these segmented data streams, two classical supervised machine learning approaches (SVM and Random Forest) were undertaken for correct prediction of seven different pedestrian activity labels. Furthermore, we implemented a deep learning classifier (CNN) for direct activity recognition using raw data. The training and testing procedure includes three types of systems: single-subject, all-subject, and leave-one-subject-out models. For performance evaluation, we used the F-score metric, which can reach up to 92.3%, 98.1% and 97.2% for these three models, respectively. CNN with raw data provides much better accuracy than the classical machine learning models. With the capability to identify pedestrian activity and thus distracted pedestrians with great accuracy, our approach lays the foundation for a smartphone application based real time P2V warning system. In this system, the vehicle's driver gets a warning in his smartphone about the nearby presence of a distracted pedestrian.

Complete street systems integrate a wide range of users in the same space, with unequal risks and responsibilities. This makes driver attention a critical factor in assuring the safety of vulnerable users. The Conditioned Anticipation of People psychological model of driver attention proposes that drivers reflexively reengage their metacognitive processes when they anticipate visually interacting with the human face or form due to the neurological priority that the brain places on human recognition. To test this model, an eye-tracking tabulation was generated from the SHRP2 Naturalistic Driving Study that measured midsegment percent of time on-task and multitasking behavior for 200 sites in Tampa, Florida and Seattle, Washington. This attention data was statistically analyzed for the impacts of a wide range of context variables using single variable ANOVA and various multivariate models such as ordered probit fractional split and ordered probit models. Context features with a strong correlation to vulnerable user presence that support driver's visual recognition of that presence were also strongly correlated with driver attention. Features like corridor width, block length, doorway density, and sense of enclosure had the largest impact. Features that did not have an impact on the potential visual connection with street users, like lane width, right of way width, onstreet parking, functional classification, or Walkscore had no impact on driver attention or weak effect sizes, despite strong correlations with vulnerable user presence. Crash history was evaluated in conjunction with the variables most sensitive to driver attention with mixed results. Many of the features that increase the potential for drivers to see and interact with people also contribute to increases in vehicle to vehicle conflicts. A decrease in crash rate with increasing sidewalk width implies that the CAP effect can have some impact on crashes. Implications for complete streets and community design are discussed.