Quantized Feedback-enhanced Large Multimodal Model for V2X Communications

Recent advances in large multimodal models (LMMs) have inspired the integration of vehicle-to-everything (V2X) communication and diverse sensing data. One of the key challenges in LMM-enabled V2X communication is the substantial sensing feedback overhead between vehicles and roadside unit (RSU). To address this, QF-LMMV2X extracts high-level semantic features from sensing data and then quantizes the extracted feature using vector quantization, effectively reducing redundancy while preserving essential semantic information. This quantization process in QF-LMMV2X significantly minimizes communication overhead, enabling efficient V2X information exchange. From the simulation results, we demonstrate that QF-LMMV2X outperforms conventional schemes by more than 25% in terms of car accident prediction accuracy.