基于2D先验的3D目标判定算法  

A 3D object discrimination algorithm based on 2D prior

作　　者：东辉 解振宁 孙浩 陈炳兴[1] 姚立纲[1] DONG Hui;XIE Zhenning;SUN Hao;CHEN Bingxing;YAO Ligang(School of Mechanical Engineering and Automation,Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福州大学机械工程及自动化学院,福建福州350108

出　　处：《福州大学学报（自然科学版）》2023年第3期387-394,共8页Journal of Fuzhou University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62173093);福建省自然科学基金资助项目(2020J01456)。

摘　　要：提出一种基于2D先验的3D目标判定算法.首先用轻量级MobileNet网络替换经典SSD的VGG-16网络,构建出MobileNet-SSD目标检测模型;其次,通过改进网络结构,提高模型对小目标的检测能力,并引入Focal Loss函数来解决正负样本不均衡和易分样本占比较高的问题;在相同数据集上,将改进算法与Faster R-CNN、 YOLOv3及MobileNet-SSD进行对比测试,其平均精度mAP分别提高了7.2%、 8.8%和10.6%;最后,通过改进算法获取ROI,利用深度相机将二维ROI转换为ROI点云,并借助直通滤波来判断目标物体是否为真实场景物体,既省去了传统点云识别中的诸多步骤又避免了点云深度学习中三维数据集制作难度较大的问题,在识别速度和识别精度上达到了较好的平衡.A 3D object discrimination algorithm based on 2D priori is proposed.Firstly,the VGG-16 network of classic SSD is replaced by the lightweight MobileNet network to construct the Mobile Net-SSD object detection model.Secondly,the network structure is improved to enhance the detection capability of small objects,and the Focal Loss function is introduced to address the issues of imbalanced positive and negative samples and high percentage of easy samples.The improved algorithm is tested again Faster R-CNN,YOLOv3 and MobileNet-SSD on the same dataset,and the mAP increased by 7.2%,8.8%and 10.6%,respectively.Finally,the ROI is obtained by the improved algorithm,and the 2D ROI is converted into a ROI point cloud by the depth camera,then using PassThrough filter to judge whether the goal object is a real scene object,which eliminates the steps of traditional point cloud recognition and avoids the difficulty of making 3D dataset in point cloud deep learning,achieves a good balance between recognition speed and recognition accuracy.

关 键 词：点云识别 MobileNet网络 SSD 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]