Radar Point Cloud-Based Deep Learning Approach for High-Capacity Urban V2I Communications

Abstract

With the evolution of 5G and 6G communication systems, sensing-assisted communication technology becomes essential for improving the operational efficiency of vehicle-to-infrastructure (V2I) systems. This letter proposes a radar point cloud-based deep learning (PDL) flow to achieve higher channel capacity between vehicles and base stations (BSs). Firstly, target-clutter segmentation is conducted using the proposed PDL based on the radar point clouds. With the segmentation results, a target clustering and tracking module is applied for a further step of false vehicle trajectory elimination. Finally, the deep learning-based method is deployed for intelligent communication beam classification of vehicles with the input of associated radar point cloud features. Numerical results show that the proposed approach improves channel capacity by approximately 10% over the deep neural network (DNN) and 2% over the DNN-based long short-term memory (LSTM) network. Meanwhile, it reduces the parameter count by about 85% compared to DNN and by 88% compared to LSTM-DNN, respectively.