Deep Learning for Multi-Modal Sensor Fusion and CSI Compression in Vehicular Communications

Vehicle-to-infrastructure (V2I) communications requires accurate channel state information (CSI) estimation with strict latency constraints, which is challenging in dynamic vehicular environments. This paper introduces a novel multi-modal fusion framework that leverages radar, camera, and GPS data to enhance CSI compression in V2I networks. We design a distributed deep learning architecture comprising an encoder at the vehicle, a sensor fusion network at the base station (BS), and a decoder that jointly optimizes CSI reconstruction under rate-limited feedback constraints. Our key innovations include: (1) systematic integration of three complementary sensing modalities, quantifying their individual and combined contributions; (2) a temporal correlation model that leverages both multi-modal data and historical CSI estimates; and (3) a comprehensive evaluation framework using our newly introduced DeepVerse 6G dataset. Extensive experiments demonstrate significant performance improvements, e.g., GPS data alone enhances CSI estimates by approximately 67% in static scenarios and 69% in dynamic scenarios compared to conventional approaches. Our analysis across varying traffic densities provides valuable design guidelines for practical deployment in diverse vehicular environments.