TSTC: Task-Supervised Temporal Semantic Communication for Vehicle-Infrastructure Cooperative Perception

Abstract

Vehicle-infrastructure cooperative perception enhances scene awareness by incorporating complementary observations from roadside units. However, delivering all roadside data is difficult under limited bandwidth, packet loss, and transmission delay. To address this issue, this paper proposes TSTC, a task-supervised temporal semantic communication framework for cooperative temporal perception. TSTC generates semantic packets from V2X-Seq-SPD annotations, estimates the contribution of each packet to perception quality, and selects informative roadside content under communication constraints. To mitigate the influence of packet loss and delay, a temporal semantic memory with decay is introduced to reuse historical context. Experiments on V2X-Seq-SPD compare TSTC with random, nearest-object, and rule-based approaches under different bandwidth limits, loss rates, and delay conditions. The results show that TSTC consistently achieves better recall and semantic precision across different communication settings while maintaining stable perception performance under constrained transmission resources.