PROJECT TITLE :

LIDAR and Position-Aided mmWave Beam Selection With Non-Local CNNs and Curriculum Training

ABSTRACT:

Due to the narrow mmWave beamwidth and high user mobility, efficient millimeter wave (mmWave) beam selection in vehicle-to-infrastructure (V2I) communication is a crucial but challenging task. V2I stands for vehicle-to-infrastructure communication. Contextual information from light detection and ranging (LIDAR) sensors mounted on vehicles has been leveraged by data-driven methods to produce useful side information. This has been done in an effort to reduce the search overhead associated with iterative beam discovery procedures. In this paper, we propose a lightweight neural network (NN) architecture along with the corresponding LIDAR preprocessing, both of which significantly outperform other works that have been done in this area. Our solution incorporates a number of novelties, each of which contributes to an increase in the model's final accuracy as well as its speed of convergence. In particular, we define a novel loss function that is influenced by the idea of knowledge distillation, introduce a curriculum training approach that exploits line-of-sight (LOS)/non-line-of-sight (NLOS) information, and propose a non-local attention module to improve the performance for the more difficult non-line-of-sight cases. All of these contributions are based on the knowledge distillation idea. Simulation results on benchmark datasets show that our NN-based beam selection scheme can achieve 79.9% throughput of an exhaustive beam sweeping approach without any beam search overhead by utilizing only LIDAR data and the receiver position. Furthermore, it can achieve 95% throughput by searching among as few as six beams. When compared with the inverse fingerprinting and hierarchical beam selection schemes, our proposed method significantly shortens the amount of time spent searching for beams in the typical mmWave V2I scenario, which is necessary in order to achieve the throughput that is desired.