Open Collections

Abstract

Autonomous driving technology has seen rapid advancements, transforming mobility and logistics. Companies such as Tesla, Waymo, and others have pioneered innovations in vehicle automation, employing diverse sensor modalities and AI-driven strategies to enhance safety and efficiency. Despite progress, challenges persist in isolated decision-making and limited situational awareness of self-driving vehicles. This research addresses these limitations by focusing on Connected and Automated Vehicles (CAV) that utilize Vehicle-to-Everything (V2X) communication, enabling dynamic interaction between vehicles, infrastructure, and vulnerable road users. This study introduces a novel cooperative perception framework leveraging multi-modal sensor fusion for improved detection of occluded and small objects, particularly pedestrians and cyclists. Additionally, federated cooperative learning methods are proposed to enable continuous, privacy-preserving training of deep neural networks for object detection, mitigating challenges of data heterogeneity and limited resources. A parameter-efficient approach reduces communication overhead, enhancing scalability and real-world applicability. The integration of Roadside Units (RSU) further bridges sensing gaps, improving safety and situational awareness in high-risk environments. To advance decision-making, this work incorporates Vision-Language Models (VLM) and Large Language Models (LLM) for contextual reasoning in autonomous driving. A two-phase, multi-vehicle framework facilitates collaborative decision-making, combining local reasoning with real-time negotiation to optimize behavior in complex scenarios like intersections and highway merging. Evaluation results demonstrate significant improvements in detection accuracy, decision quality, and system efficiency, highlighting the potential of cooperative and connected autonomy to revolutionize transportation systems. This research sets a new standard for scalable, adaptable, and intelligent autonomous driving technologies.