PillarFocusNet: 3D Point Cloud Object Detection Network

This study focuses on 3D point cloud object detection. PillarFocusNet, a 3D point cloud object detection network was proposed in this study by improving and optimizing the PointPillars. In this research, the conventional random dropout technique was substituted with the DBSCAN density clustering method to tackle the issue of uneven distribution in 3D point cloud data. Additionally, a Mixed Pooling Dilated Convolution Layer (MPDC layer) was developed to address the sparsity of the point cloud data. Experiments on the KITTI dataset demonstrate that PillarFocusNet enhances bbox, BEV, and 3D detection accuracy by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.1 {\%}, 2.0 {\%}$</tex>, and 1.3 %, respectively, when compared to PointPillars.