Semantic and Risk-Aware Handoff Coordination for Cooperative Perception in Highway V2XNetworks with Strict Tail-Latency Requirements

Abstract On highways, cooperative perception depends on connected vehicles sharing urgent descriptions of nearby scenes as they pass through overlapping coverage from neighbouring next-generation base stations (gNBs). During that short interval, perception messages and handoff signalling draw on the same limited pool of radio resources. Current group-handoff methods improve association, balance network load, or reduce average delay, yet usually assume a fixed perception payload and do not control the rare delay spikes that matter most for safety. We introduce Semantic Risk-Aware Handoff Orchestration (SRAHO), a cross-layer controller that makes three decisions together: it chooses the smallest semantic payload that meets a required perception-retention level; selects direct vehicle-to-infrastructure (V2I) transmission or handoff-agent (HOA) relaying while accounting for mobility, fading, and the possibility that a helper leaves; and assigns resource blocks (RBs) to minimise the Conditional Value-at-Risk (CVaR) of handoff delay. Our analysis derives closed-form rate moments and high- and low-SNR limits for Rayleigh fading, exact upper and lower bounds for fading-truncated service moments and their quantile-grid scaling, and a Gaussian-margin test for the persistence of a V2V link. It also proves strict convexity of M/G/1 handoff delay with respect to RB allocation, gives a square-root spare-capacity rule for mean-optimal allocation, and obtains a closed-form Lundberg-type upper bound linking queueing-delay CVaR to its mean. For separable discretely convex objectives, the CVaR-greedy allocator is exact; a computable a posteriori certificate bounds any remaining optimality gap. SRAHO delivers the lowest CVaR0.95 handoff delay in all eight stress scenarios. Under interference, it cuts CVaR0.95 from 60.01 ms with Direct V2I to 51.18 ms, a 14.72% reduction, while handoff success rises from 0.955 to 0.983. Semantic adaptation also reduces the payload per step from 100 608 to 69 406 bits, or 31.0%. In model-based tests under interference, CVaR-based allocation reduces deadline violations by 29% against mean-delay allocation. The greedy method matches exhaustive optimization in 1198 of 1200 cases, and the full decision for 12 vehicles takes 8.85 ms.