ORCID

0000-0002-8660-5240

Keywords

Connected and Autonomous Vehicles, DNN-based Vehicular Trajectory Prediction, VANETs, Model Based Communication, Multi-hop Vehicular Communication

Abstract

This dissertation presents a unified framework for communication-aware trajectory prediction and vehicle state tracking in connected and autonomous vehicle (CAV) systems. Unlike most deep learning–based forecasting models that assume full observability of surrounding agents, the proposed approach explicitly accounts for limited communication range, packet loss, and bandwidth constraints that characterize real vehicular networks.

A model-based, error-driven information-passing policy is introduced to extend situational awareness beyond one hop. Rather than relaying messages based on distance or freshness, each vehicle evaluates the utility of received content in reducing system-wide position tracking error (PTE). Using compact Gaussian Mixture Model (GMM) representations of motion intent, vehicles rank and rebroadcast neighboring agents’ models according to their contribution to tracking accuracy. These parameters are transmitted within the optional fields of SAE J2735 Basic Safety Messages (BSM Part II), enabling a communication-efficient, semantics-aware multi-hop strategy that remains robust under varying network load. \sloppy On this infrastructure, two deep probabilistic forecasting models are proposed: the Context-Aware Gaussian Mixture Model (CA-GMM) and the Context-Aware Attention-based GMM (CAA-GMM). CA-GMM fuses motion history and rasterized scene context to generate multimodal trajectory distributions, while CAA-GMM enhances this with a transformer-based attention mechanism that captures inter-agent interactions. Both models align with the model-based communication paradigm, enabling compact, uncertainty-aware transmission of predictive states.

Evaluated on the nuScenes and Argoverse 2 benchmarks and tested under realistic communication and observation loss, the proposed models achieve competitive multimodal forecasting accuracy. The attention-enhanced CAA-GMM further improves interaction modeling while preserving efficiency, establishing a scalable and interpretable foundation for communication-constrained cooperative autonomy.

Completion Date

2025

Semester

Fall

Committee Chair

Pourmohammadi Fallah, Yaser

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Electrical and Computer Engineering

Format

PDF

Identifier

DP0029746

Document Type

Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS