CogniTensor: Adaptive Multi-Sensor Transmission for V2X Collective Perception via Uncertainty and Channel Awareness

Abstract

Collective Perception (CP) in Vehicle-to-Everything (V2X) networks extends vehicle perception, but in high-density scenarios, simultaneous multi-sensor data exchange saturates the available 5.9 GHz channel bandwidth, degrading the Age of Information (AoI) and compromising safety. Existing approaches employ static transmission strategies that disregard both the semantic relevance of perceived data and real-time channel dynamics, leading to suboptimal resource utilization. CogniTensor addresses this gap by formulating adaptive multi-sensor transmission as a Markov Decision Process (MDP) solved via a Discrete Soft Actor-Critic (SAC-D) agent. It is the first framework to jointly optimize sensor selection, adaptive Tucker-based tensor compression, and transmission timing by reasoning over a composite state that includes perceptual uncertainty from Evidential Deep Learning, channel conditions, and vehicle dynamics. Evaluation demonstrates that CogniTensor achieves 58% lower bandwidth consumption and 9% higher resource efficiency (R/BW: 1.29 vs. 1.18) relative to static full-sensor transmission, while maintaining competitive AoI through intelligent, congestion-aware sensor selection.