Choosing an Information Representation for Cooperative Perception

Determine which information representation scheme should be adopted for cooperative perception in vehicle-to-everything (V2X) systems in practical traffic environments, balancing bandwidth and latency constraints, minimizing information loss, and enabling standardization across heterogeneous agents; specifically, decide among data-level (raw sensor data exchange), feature-level (learned intermediate features), and object-level (high-level detections) representations or design an alternative that satisfies these objectives.

Background

The survey categorizes cooperative perception message types into data-level, feature-level, and object-level representations. Data-level sharing preserves full fidelity but is impractical due to extreme bandwidth demands. Feature-level sharing reduces volume but faces challenges in standardization and heterogeneity because learned features are model-dependent. Object-level sharing is bandwidth-efficient and interpretable but loses rich information needed for tasks beyond object detection.

Given these trade-offs, no single approach is universally optimal across communication efficiency, information retention, and standardization. The authors highlight the need for a representation choice suitable for real-world deployments and suggest possible directions such as explicit compact 3D geometric representations and unified feature spaces.