F-PCNet: A New Fast Object Detection Method Based on Point Cloud Only*

Abstract

Although deep learning methods have greatly improved the accuracy of the object detection tasks, it is still challenging to balance the efficiency and accuracy of the algorithms under circumstances of point clouds only. In this paper, an anchor-free one-stage deep neural network, F-PCNet, is proposed to realize real-time detection based on point clouds on an autonomous driving platform while maintaining high accuracy. The proposed network takes the bird’s eye view of point clouds collected by LiDAR as input, and outputs the category and 2D bounding box of each detected object. The backbone of F-PCNet is composed of residual network modules of different sizes which effectively reduce the impact of learning degradation. The anchor-free detection head enables F-PCNet to achieve high levels of accuracy and efficiency. Experimental results show that F-PCNet achieves high detection accuracy in a short time consumption and is suitable for real-time detecting scenarios.