Deep Learning for Ego- and Scene-Centric Vehicle Trajectory Prediction at Cooperative Intersections for Vulnerable Road Users Safety

Abstract

Intersections are highly dynamic environments where multiple traffic flows, diverse road users, and visibility obstructions create safety challenges. Adopting Autonomous Vehicles (AVs) adds further complexity, particularly in mixed-traffic scenarios where AVs and human-driven vehicles interact, introducing uncertainty. To address these challenges, this dissertation explores the integration of deep learning-based trajectory prediction with V2X communication to enhance predictive safety applications at cooperative intersections. By leveraging infrastructure-recorded data beyond the perception limits of onboard sensors, V2X communication enables more informed and accurate trajectory forecasting, contributing to proactive collision avoidance. The key contributions of this research are the development of Fisheye-MARC, a trajectory dataset collected using a fisheye camera at an urban intersection, and multiple deep learning-based trajectory prediction networks for increasing road users' safety. This dataset facilitates the training and evaluating deep learning models for vehicle trajectory prediction in real-world traffic conditions. We investigate a predictive safety alert system structure and three primary prediction approaches: LSTM-based, ego-centric, and scene-centric. While the first two model road agents individually, the scene-centric approach holistically captures interactions among all traffic participants. To improve prediction accuracy and efficiency, we introduce the Heterogeneous Decision-Aware Attention Graph Transformer (HDAAGT). This scene-centric trajectory prediction model employs a Decision-Aware Attention Graph (DAAG) architecture to capture spatial dependencies among traffic agents. HDAAGT integrates multiple contextual inputs, including lane geometry, traffic light status, and vehicle interactions, achieving an Average Displacement Error (ADE) of 9.85 pixels (65 cm) in real-world coordinates. Comparative analysis demonstrates that the scene-centric approach is both computationally efficient and scalable, significantly reducing redundant computations by collectively processing the entire traffic scene rather than per-agent inference. Despite these advancements, several challenges must be addressed for real-world deployment. Refining the adjacency matrix in HDAAGT could further filter out non-influential agents, enhancing inference speed and model efficiency. Additionally, optimizing the model for edge deployment and integrating driver behavior modeling into V2X-enabled trajectory prediction could further improve accuracy and robustness. This research underscores the transformative potential of V2X-enabled deep learning models in predictive safety applications, paving the way for intelligent and cooperative traffic systems that enhance road safety.