Deep Learning-Based Cramér-Rao Bound Optimization for Integrated Sensing and Communication in Vehicular Networks

Integrated sensing and communication (ISAC) is capable of achieving both heterogeneous connectivity and highly accurate sensing performance in vehicular networks through effective beamforming design at the roadside unit (RSU). In the traditional paradigm, the first step is predicting the kinematic parameters of each vehicle and then designing the optimal beamforming matrix, which requires excessively large computational complexity. To tackle this issue, this paper proposes a deep learning (DL)-based method that bypasses explicit channel estimation and directly optimizes beamformers to minimize the Cramér-Rao Bound (CRB) of radar sensing while guaranteeing an acceptable level of achievable communication rate. This is achieved by leveraging the convolutional and long short-term memory (CLSTM) neural networks to implicitly capture the features of historical channels, thereby improving the ISAC system performance. Finally, simulation results demonstrate that the proposed approach can satisfy the pre-defined requirement of achievable rate, while simultaneously achieving sensing performance that approaches the perfect beamforming bound.