Mask-SL RCNN: Feature-Enhanced 3D Object Detection Network for Point Clouds

Abstract

At present, with the original point cloud as input, most of the object detectors use Pointnet++ to extract features of the point cloud based on the Farthest Point Sampling (FPS). However, affected by FPS, feature extraction is incomplete and unstable. Moreover, high-level semantic features lack the internal vertex properties of Regions of Interest (RoI). In order to solve the above problems, we propose the Mask-SL RCNN (Mask-Spherical-neighborhood-global-feature-Layer Region-CNN), a feature-enhanced 3D object detection network. It improves sampling of the farthest point through point-level feature enhancement. In addition, we propose Spherical neighborhood global feature Layer (SL) to supplement the global features and improve the learning ability of network. At last, based on semantic-level feature enhancement, we design grid pooling layer based on vertex attention, which fully explores the boundary characteristics of RoI and increases ability to learn advanced features in RoI. Our network improves detection precision of small objects such as pedestrians. Compared with PointRCNN, it has improved the mAP of simple, medium, and difficult object detection in the KITTI dataset by 2.66%, 1.69%, and 0.67%, respectively.

Keywords

• 3D object detection
• LiDAR point clouds
• convolutional neural networks