A Predictive-Prescriptive Safety Framework at Intersections in a Connected Vehicle Environment

Abstract

The connected and automated vehicle (CAV) technology in recent years has demonstrated its potential in improving efficiency in transportation systems. Prediction, as a key component of the technology, enables smart vehicles to anticipate future movements of traffic agents and potential future risks, so as to plan in advance by incorporating these predictions in their trajectory planning. In this dissertation I propose a prediction-based framework to identify risky scenarios at urban intersections, develop strategies to mitigate them, and conduct prediction-based vehicle trajectory planning in a connected environment. The framework consists of three main components: (1) real-time risky driving prediction; (2) traffic agent trajectory prediction; (3) prediction-based vehicle trajectory planning. For risky driving prediction, I propose an unsupervised learning framework to predict risky driving at urban intersections in a connected vehicle environment. The proposed framework uses time series k-means to categorize multi-dimensional time series trajectories into several context-aware driving patterns. I train an anomaly detection model on the trajectory dataset to identify anomalous trajectories, and apply this model to clusters of driving patterns to provide Risky Driving Prediction(RDP) scores for each driving pattern. I provide a real-time online assessment approach to predict the risk score of driving trajectories that travel toward a signalized intersection. For pedestrian trajectory prediction, I evaluate the model using a benchmark dataset and a proprietary dataset collected at urban intersections.I compare the performance of step attention with three existing state-of-the-art algorithms. Experiments show that on the benchmark dataset the average and final displacement errors of step attention for a 4.8-seconds prediction horizon are 0.53 and 1.72 meters, respectively. Both average and final displacement errors are favorable compared to the benchmark methods. Furthermore, on the urban intersection dataset, the proposed model has an average displacement error of 0.74 meters and a final displacement error of about 1.40 meters for a 6-seconds prediction horizon. I conduct a complementary set of experiments to further investigate model performance in a real-world intersection. In these experiments, the model gains an ADE/FDE of 0.76/1.70 m. The proposed model also produces accurate prediction results on different scenarios composed of different walking patterns (e.g., straight and curvy) and different environments (e.g., sidewalk and street). Finally, I extend the architecture to different agents, namely, vehicles and cyclist, and demonstrate that step attention can perform well on long look-ahead trajectory prediction tasks for heterogeneous agents. With aforementioned models and results, I put forward a predictive trajectory planning framework to help autonomous vehicles plan future trajectories. I develop a partially observable Markov decision process (POMDP) to model this sequential decision making problem, and a deep reinforcement learning solution methodology to learn high-quality policies. The POMDP model utilizes driving scenarios, condensed into graphs, as inputs, to devise safe, comfortable, and energy-efficient trajectories for the subject vehicle to follow. I propose a simulation framework to generate socially acceptable driving scenarios using a real world autonomous vehicle dataset. I evaluate the proposed work in two complex urban driving environments: a non-signalized T-junction and a non-signalized lane merge intersection. The framework demonstrates promising performance for 5 seconds planning horizons. I compare safety, comfort, and energy efficiency of the planned trajectories against human-driven trajectories in both experimental driving environments, and demonstrate that it outperforms human-driven trajectories in a statistically significant fashion in all aspects.