基于PointPillars改进的点云目标检测算法  

Improved Point Cloud Object Detection Algorithm Based on PointPillars

作　　者：何俊杰 任明武[1] HE Junjie;REN Mingwu(School of Computer Science and Engineering,Nanjing University of Science&Technology,Nanjing 210094)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机与数字工程》2023年第10期2313-2317,共5页Computer & Digital Engineering

摘　　要：现有的点云目标检测方法虽然层出不穷,但主要以提升检测精度为主,检测效率较低,很难满足实际应用时的实时性需求。PointPillars[1]方法通过一种创新的点云特征处理方法,明显提升了点云目标检测的速度,为工业界的实际应用提供了可能性,但其精度相比速度略有欠缺。论文基于多尺度特征融合和3D注意力机制,对PointPillars方法的结构进行改进,提高了原方法对于多尺度目标的检测精度,弥补了该方法相比主流方法在检测精度上的不足。改进后的方法兼具较高的检测精度和较快的检测速度,具有一定的工程应用价值。Although the existing point cloud object detection methods are constantly being proposed,these methods mainly focus on improving the detection accuracy,but their efficiency of detection is not high enough,it is difficult to meet the real-time requirements in practical applications.The PointPillars method significantly improves the speed of point cloud object detection through an innovative point cloud feature processing method,which provides the possibility for practical applications in the industry.Combining the idea of multi-scale feature fusion and 3-D attention mechanism,this paper improves the structure of PointPillars method to improve the detection performance of the original method for multi-scale targets,which makes up for the lack of detection accuracy of this method compared with state-of-the-art methods.The improved method has better detection performance still with enough detection speed,and has certain value of engineering application.

关 键 词：点云 目标检测 PointPillars 多尺度 3D注意力 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]