Graph signal processing based object classification for automotive RADAR point clouds

Abstract

As the automotive industry evolves, visual perception systems to provide awareness of surroundings to autonomous vehicles have become vital. Conventional deep neural networks have been effective on 2D Euclidean problems during the previous decade. However, analyzing point clouds, particularly RADAR data, is not well-studied due to their irregular structures and geometry, which are unsuitable for 2D signal processing. To this end, we propose graph signal processing (GSP) based classification methods for RADAR point clouds. GSP is designed to process spatially irregular signals and can directly create feature vectors from graphs. To validate our proposed methods experimentally, publicly available nuScenes and RadarScenes point cloud datasets are used in our study. Extensive experiments on these challenging benchmarks show that our proposed approaches outperform state-of-the-art baselines.