基于点云深度学习的3D目标检测  

3D object detection based on point cloud deep learning

作　　者：敖建锋 苏泽锴 刘传立[1] 李美妮 朱滨 AO Jian-feng;SU Ze-kai;LIU Chuan-li;LI Mei-ni;ZHU Bin(School of Architectural and Surveying Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学建筑与测绘工程学院,江西赣州341000

出　　处：《激光与红外》2020年第10期1276-1282,共7页Laser & Infrared

基　　金：国家自然科学基金地区基金资助项目(No.41561091);江西省教育厅科学技术研究项目(No.GJJ150663);江西省教育厅科学技术研究项目(No.GJJ150629);江西省教育厅科学技术研究项目(No.GJJ180501)资助。

摘　　要：针对目前利用点云进行3D目标检测的研究较少和检测精度不高的问题,利用Frustum-Pointnets模型实现基于点云的3D目标检测,并在该模型的基础上进行改进,选用不同的激活函数和参数初始化方法进行组合对比,进一步提高模型的精度。实验表明:在选用Swish激活函数和He参数初始化方法时汽车平均检测精度提高了0.31%,行人平均检测精度提高了0.41%,骑车人平均检测精度提高了5.5%。因此改进后的模型能有效提高检测的精度,使得模型能够应用在复杂的场景中。In terms of current situation of insufficient researches and accuracy issue for 3D Object Detection of Point Cloud,higher model accuracy is possible to be achieved by improved Frustum-Pointnets model based on 3D Object Detection of Point Cloud,and under the cross-reference by means of different activation and initialization.Experimental results shows that using Swish activation and He initialization for detecting bring about,averagely,0.31%accuracy growth for cars detection,0.41%for pedestrians and 5.5%for cyclist.Accordingly,improved model is able to enhance the detection accuracy to be applied to complex occasions.

关 键 词：点云 深度学习 3D目标检测 Frustum-Pointnets 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]