GCN-Net: 3D Point Cloud Classification & Localization Using Graph-CNN

Abstract

This study aims to demonstrate the application of a
graph convolutional neural network for the purpose of object
detection in a LiDAR point cloud. To achieve efficient encoding of
the point cloud, the research used a near-neighbours graph with a
predetermined radius. The study created a graph convolutional
neural network so that study can find out what kind of object and
what class each vertex in a graph actually represents. The research
developed a box merging and scoring operation to reliably
integrate detections from multiple vertices into a single score, as
well as an auto-registration technique to reduce the amount of
internal translation errors. Based on the findings obtained from
our experimentation using the KITTI benchmark, the research
has arrived at the inference that the proposed technique
demonstrates a commendable level of precision when compared to
the point cloud. In fact, in some cases, it outperforms fusion-based
methods. Based on the findings of our study, it can be concluded
that the graph neural network has promising potential as a novel
and efficient tool for the identification and recognition of threedimensional objects