Multi-agent Collaborative Perception for Autonomous Driving: Unsettled Aspects

This report delves into the field of multi-agent collaborative perception (MCP) for autonomous driving: an area that remains unresolved. Current single-agent perception systems suffer from limitations, such as occlusion and sparse sensor observation at a far distance.

Multi-agent Collaborative Perception for Autonomous Driving: Unsettled Aspects addresses three unsettled topics that demand immediate attention:

• Establishing normative communication protocols to facilitate seamless information sharing among vehicles
• Defining collaboration strategies, including identifying specific collaboration projects, partners, and content, as well as establishing the integration mechanism
• Collecting sufficient data for MCP model training, including capturing diverse modal data and labeling various downstream tasks as accurately as possible