In recent years, shared spaces are gaining popularity in urban planning and transportation engineering, which is encouraging the reduction in motor vehicle dependency and promoting shared street design. This approach increases the interaction between motorized and non-motorized road users, which mandates the understanding of their interactions to design safe, sustainable, and equitable shared spaces. To facilitate the understanding of these interactions between road users, it is important to have a well-developed trajectory prediction and generation model. Considering this motivation, this thesis aims to develop a trajectory generation model using deep reinforcement learning that considers social context as observation. The first part of this thesis focuses on pedestrian trajectory generation using social context in a vehicle crowd interaction space, while the second part extends the study to encompass both pedestrian and bicycle trajectories. This method explores different social observation and action generation methodologies for different road users. The developed model is validated using two benchmark datasets DUT and SDD dataset. Average Displacement Error and Final Displacement Error are used to assess the performance of these models. In the final part of the thesis, an early conflict detection system is developed with the application of trajectory generation. The developed pedestrian trajectory generation model can be effectively used in simulation environments. The contribution of this thesis is using social context to generate observation and using action clusters to discretize action space for reinforcement learning agents. It also combines trajectory generation, deep learning, and early conflict detection to assist urban planners, transportation engineers, and shared space designers. The outcome of this research provides valuable insights to make shared spaces safer, more efficient, and more sustainable.

Longitudinal pavement markings play a significant role on the roadways by delivering information to motorists to help them navigate and follow the road. These markings are also considered to be a crucial control device that can enhance ideal nighttime visibility, especially on rural roads where the surrounding luminance is insufficient. Hence, the main question for public agencies or officials is about when the replacement of the pavement markings needs to take place. The Federal Highway Administration (FHWA) is considering proposing a minimum level of retroreflectivity standard and based on that, the Manual on Uniform Traffic Control Devices (MUTCD) set aside a draft (FHWA 2010). Currently, there are no specific guidelines on maintaining the pavement marking retroreflectivity. The purpose of this research is to develop a pavement marking degradation model using machine learning techniques, which is known as the most powerful technique nowadays. Using such a technique has become more popular than ever with respect to massive growth and the variety of available data to produce complex models with high accuracy compared to traditional statistical methods. In this study, we compared four different algorithms: Support Victor Machine (SVM), Gradient Boosting (GB), Random Forest (RF), and Decision Tree (DT). This study gives the performance measurements of those algorithms using three scenarios. The outcomes demonstrate that all of the algorithms perform magnificently, with accuracy ranging between 87 and 90 percent and an acceptable level of efficiency for Recall (78 percent).

Crash severity models are typically developed using police reported injury severity databases. However, several research studies have identified various challenges associated with police reported data. Therefore, the current dissertation is focusing on developing high resolution crash severity models based on medical professional driver injury severity reported using Abbreviated Injury Scale for eight body regions. The dissertation focused on developing a disaggregate injury severity modeling framework that can enhance the estimation accuracy of independent variable impacts on severity. Within this broad research vision, the dissertation has multiple objectives. First, a joint random parameters multivariate model structure with as many dimensions as severity by body location was developed. The empirical analysis involves the estimation of Random Parameters Multivariate Generalized Ordered Probit Model that allows for the influence of common unobserved factors affecting the vehicle occupant severity across body locations. Second, we incorporate the influence of temporal factors (observed and unobserved) within a multivariate model system for medical professional generated body region specific injury severity score. For this purpose, we adopt a hybrid econometric modeling approach that accommodates for the unobserved factors. Third, the dissertation compares the predictive performance of the state-of-the-art econometric model with the predictive performance of state-of-the-art machine learning methods. We consider machine learning approaches such as Random Forest, Logistic Regression, Boosting, and Support Vector Machine. Finally, the dissertation applied two approaches. First, analyze dependent variables with an ordered logit framework using the six injury severity levels. The second approach is adopting a hurdle ordered logit framework by splitting the dependent variable into two stages: binary and truncated which exclude the zero cases. The model performance of these approaches are compared using the data of two body regions.

Time-specific Safety Performance Functions (SPFs) were proposed to achieve accurate and dynamic crash frequency predictions and bridge the gap between annual crash frequency prediction and real-time crash likelihood prediction. This research proposed time-specific SPFs considering the temporal variation in crashes and traffic characteristics. Firstly, the developed time-specific SPFs that include different ATM strategies (i.e., HOV, merge, diverge and reversible lanes segments) were investigated in this study. The results indicate that the traffic turbulence during specific hours would relate to crash occurrence. Further, the variables that represent the speed and occupancy differences between the HOV lanes/reversible lanes and general-purpose lanes were found to be positively associated with crash frequency. Moreover, the design of the reversible lane segments, the number of access points positively impacts the crash frequency. Secondly, this study proposed different methodologies to improve the prediction accuracy of time-specific SPFs. The model comparison including the negative binomial model, Poisson lognormal model and hierarchical Poisson lognormal model. The results showed that the Poisson lognormal model outperformed the negative binomial model. Moreover, the hierarchical models outperformed the corresponding Poisson lognormal model. Other than prediction accuracy, this study also successfully identified the factors associated with the different crash types or severity in crash frequency prediction models. Finally, this study proposed a novel iterative imputation method to impute the 100% missing volume and speed data from the different states with similar crash rates. The crash rates for 18 states were calculated and the ANOVA test was applied to group the states with similar crash rates. Afterward, this study used FL and VA, which both have traffic data to test the proposed iterative imputation method. The results indicated that the imputed traffic data could capture the same traffic pattern as the real-collected traffic data. Further, The MAE between the imputed volume and the real-collected volume for FL is 2.47 vehicles/3hrs/segment. The MAPE between the imputed and real-collected volumes for FL is 11.07%. Moreover, this study applied the proposed iterative imputation method to develop time-specific SPFs for the state without traffic data and compared the results. The results illustrated that the time-specific SPFs developed by imputed traffic data perfectly reflected the significant variables for both morning and afternoon peak models, with a prediction accuracy of 87.1% for the morning peak model. This could help the traffic operators in the states without high-resolution traffic data to determine the factors contributing to crash occurrence on freeway segments during a specific time period.

Thesis (M. Arch.)--University of Hong Kong, 2002. / Includes bibliographical references.

Thesis (M.Arch.)--University of Hong Kong, 2002. / Includes bibliographical references. Also available in print.

Thesis (M.S.)--Lehigh University, 2009. / Adviser: James E. Roberts.

<p> A method for accurate emissions estimation that will contribute to promoting public health has been increasingly important. The purpose of this study is to develop a novel method that is designed to make accurate real-time emissions estimation from individual vehicles on freeways possible. The benefit of this method is that it can overcome the weakness of macroscopic emissions estimation methods, which underestimated emissions. </p><p> The most distinguishing feature of the Speed Profile Estimation (SPE) method is that it uses a speed profile (SP) that is generated by the sum of a basic SP (BSP), which is calculated by the basic travel information of an individual vehicle obtained from vehicle reidentification (REID), and a residual SP (RSP), which is estimated by categorized traffic information. </p><p> In order to estimate RSP this research employs Autoregressive (AR) model and Fourier series (FS). And to find the parameters of RSP, the total absolute difference between actual SP emissions and estimated SP emissions was optimized by genetic algorithm. For this, parameters are calculated for all possible combinations of three categorizations and clusters by K-mean clustering. Individual vehicle trajectories from two freeways, US101 and I-80, were provided by the Next Generation Simulation (NGSIM) dataset. US101 was examined for calibration, and I-80 for validation. And then, transferability tests were conducted for various section distances to verify model transferability. Finally, REID is simulated with low vehicle signatures match rates to test its applicability to real situations. </p><p> Unlike previous methods, the SPE is notable for its real-time, transferable, reliable, and cost efficient emissions estimation. The calibration and validation account only 4.0 % and 4.1 % MAPEs, respectively. Moreover, transferability tests showed that MAPEs are lower than 4.4 % in both longer and shorter section distances. Furthermore, REID simulation increases only 0.2 % MAPE even in low vehicle signatures match rates, which is lower than 5 % MAPE in emissions estimation. </p><p> Any signal-like formulation other than AR or FS can perform better emissions estimation when it replaces the RSP. Also, in this research the SPE method was calibrated only for LOS F, when it is arguably of greatest value, but further research should be coordinated to extend the models in other possible traffic conditions such as LOS A~E.</p>

Thesis (M.App.Sc.) - Carleton University, 2002. / Includes bibliographical references (p. 176-185). Also available in electronic format on the Internet.

Evaluating the impact of longer change and clearance intervals on signalized intersections and corridors is the main goal of this study. In fact, the Florida department of Transportation (FDOT) has adopted a new signal retiming effort in a number of signalized intersections along several corridors. The Orange County started implementing the new signal timing from December, 2013 and completed it in June, 2015. The other objective of this new signal timing is to minimize the red light running rate. This study is dedicated to investigate the signal retiming effort adopted by the FDOT and how the new signal timing might impact the studied signalized intersections' performance and safety. To address this issue, a number of signalized intersections along three corridors in Orange County were investigated during different three time of the day periods AM, MD, and PM. Additionally, three categories of signal timings were adopted to better understand the performance and safety of old (pattern 1), current (pattern 2), and proposed (pattern 3) signal timings. The analysis was based on the Simtraffic simulation which is a part of Synchro 8 software. The research results provide that the signalized intersection's performance along the three corridors during the three plans of the day were found significantly affected by lengthening the change and clearance intervals. Signal timing 2 and 3 were observed significantly different than signal timing 1 which have greater intersection delay, queue length, intersection overall volume to capacity v/c ratio, and Intersection capacity utilization ICU. Furthermore, the results show that the signal timing 2 and signal timing 3 significantly increase the total delay and travel time along the studied arterials during the three plans of the day.