Bandwidth-Accuracy Tradeoffs in V2X Cooperative Perception: A Comparative Review

Abstract

Vehicle-to-Everything (V2X) cooperative perception extends the perception horizon of autonomous vehicles by enabling information exchange between vehicles and roadside infrastructure. Its practical deployment is constrained by communication bandwidth, latency, and perception accuracy, since raw sensor data often exceed realistic wireless-channel capacity. This paper presents a comparative review of V2X cooperative perception methods from the bandwidth–accuracy perspective. It provides a five-category taxonomy of fusion paradigms: early, late, intermediate feature-level, hybrid, and learned-compression fusion. Representative methods are compared by communication cost, reported AP@0.5, fusion stage, and technical principle. The review formalizes the bandwidth–accuracy tradeoff as a constrained optimization problem with latency as an additional deployment constraint. The analysis shows that raw-data fusion approaches an upper-bound accuracy region but is impractical, while late fusion is communication-efficient but accuracy-limited. Intermediate feature fusion with spatial selectivity and learned compression occupies the most favorable practical region. The paper also discusses the simulation-to-real gap and open challenges, including communication imperfections, adversarial robustness, heterogeneity, scalability, privacy, and non-standardized evaluation.