UCLA

Automated driving technology has emerged in recent years due to its potential to revolutionize transportation, bringing enhanced safety and efficiency. However, large-scale deployment is restricted by challenges inherent to single-vehicle systems, including occlusions, interactions with diverse traffic elements, and complicated decision-making. This dissertation advances the realm of Cooperative Driving Automation (CDA) as a solution, focusing onsimulation frameworks and cooperative perception algorithms design.

The research starts with introducing OpenCDA, a comprehensive simulation frameworkfor CDA system prototyping, and OPV2V, the first large-scale simulated cooperative perception dataset. These tools address the need for a simulated environment to prototype and validate CDA algorithms, bridging existing gaps in cooperative perception advancement.

Built upon OpenCDA and OPV2V, I present two state-of-the-art cooperative perceptionalgorithms. The first, a cooperative 3D LiDAR detection framework, employs a Vision Transformer architecture to tackle challenges like sensor heterogeneity, localization error, and bandwidth constraints. The second, CoBEVT, is a pioneering multi-agent, multi-camera perception framework that uses economical RGB cameras to generate Bird-eye-view map predictions, offering a cost-effective solution.

The final segment of the research emphasizes real-world deployment. I present V2V4Real, the first real-world dataset for V2V perception, detailing its comprehensive benchmarks and introducing novel tasks. Further, I delve into strategies to optimally train cooperative perception models using simulated data, introducing a novel module, the Homogeneous Training Augmenter, which demonstrates the efficacy of simulation in real-world applications.

In essence, this thesis provides significant contributions to the domain of CDA, offering tools, datasets, and algorithms that pave the way for the broader, real-world implementation of cooperative automated driving.