Learning to Sense: Efficient Semantic Transmission for Cooperative Object Detection at Intersections

To enable intelligent transportation applications of the future, e.g., with autonomous vehicles, it is crucial to detect pedestrians, cyclists, and vehicles with high accuracy. While cooperative detection by roadside infrastructure and connected vehicles jointly can provide high accuracy, it comes with an associated cost due to transmission and processing of massive multimodal sensory data, including high-dimensional images and Li-DAR point clouds, which impose significant communication and computation burdens on both vehicle and infrastructure sides. To decrease this burden without a significant loss in detection accuracy, we propose RL-VoIS, a reinforcement learning-based Value-of-Information (VoI)-aware semantic computing and transmission scheme. RL-VoIS enables vehicles to intelligently decide when and which sensory data should be semantically processed and transmitted. By jointly considering semantic importance, data freshness, temporal drift, and environmental dynamics, RL-VoIS effectively suppresses unnecessary semantic computation and transmission through selective processing of high-VoI features, optimizing the trade-off between perception accuracy and end-to-end energy consumption. Experiments on a real-world intersection dataset demonstrate that RL-VoIS achieves superior energy and communication efficiency while maintaining high perception accuracy. RL-VoIS exhibits operational stability with zero safety violations and consistently outperforms both state-ofthe-art threshold-based and learning-based baselines.