基于改进Frustum PointNet的3D目标检测  

3D Object Detection Based on Improved Frustum PointNet

作　　者：刘训华 孙韶媛[1,2] 顾立鹏 李想 Liu Xunhua;Sun Shaoyuan;Gu Lipeng;Li Xiang(College of Information Science and Technology,Donghua University,Shanghai 201620,China;Engineering Research Center of Digitized Textile&Fashion Technology,Ministry of Education,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620 [2]东华大学数字化纺织服装技术教育部工程研究中心,上海201620

出　　处：《激光与光电子学进展》2020年第20期320-326,共7页Laser & Optoelectronics Progress

基　　金：上海市科委基础研究项目(15JC1400600)。

摘　　要：提出对图像和激光雷达点云数据进行3D目标检测的改进F-PointNet(Frustum PointNet)。首先利用图像的2D目标检测模型提取目标2D区域,并将其映射到点云数据中,得到该目标的点云候选区域,然后预测候选区域的3D目标掩模,最后利用掩模对3D目标进行检测。当预测掩模时,提出的宽阈值掩模处理可以用来减少原始网络的信息损失;增加注意力机制可以获取需要被关注的点和通道层;使用Focal Loss可以解决目标与背景不平衡的问题。通过多次对比实验,证明宽阈值掩模处理可以提高3D目标检测的准确率,同时注意力机制和Focal Loss可以提高预测的准确率。An improved F-PointNet(Frustum PointNet)for 3D target detection on image and lidar point cloud data is proposed.First,the 2D target detection model of the image is used to extract 2D region of the target,and it is mapped to the point cloud data to obtain the candidate region of the target.Then,the 3D target mask of the candidate region is predicted.Finally,the 3D target is detected by using mask.When the mask is predicted,the proposed wide-threshold mask processing is used to reduce the information loss of the original network,the attention mechanism is added to obtain the points and channel layers that require attention,the Focal Loss can solve the imbalance between the target and the background problem.Through multiple comparison experiments,it is proved that wide-threshold mask processing can improve the accuracy of 3D target detection,and the attention mechanism and Focal Loss can improve the accuracy of prediction.

关 键 词：机器视觉 激光雷达 点云数据 3D目标检测 宽阈值掩模处理 

分 类 号：TN958.98[电子电信—信号与信息处理]