摘要
最远点采样(FPS)算法可用于三维(3D)目标检测算法中关键点的采集,针对FPS采集的关键点中前景比例较低的问题,提出了一种基于语义分割特征的区域卷积神经网络(Seg-RCNN)3D目标检测算法。用一种基于语义分割特征的最远点采样(SegFPS)算法预测点的语义分割类别,以提高采集关键点中前景点的比例,从而提升Seg-RCNN算法的检测精度。该算法以激光点云作为输入,在第一阶段,利用3D稀疏卷积网络和2D卷积生成候选框和前景点分割网络(SegNet),得到每个点的分割类别;在第二阶段,基于SegNet输出的分割类别用SegFPS算法从原始点云中采集一小部分关键点,从而在降低算法时间复杂度和空间复杂度的同时保留一定比例的前景点和背景点。在KITTI测试集上的测试结果表明,相比现有的主流算法,Seg-RCNN算法的检测精度高、运行时间短,对中等级、容易等级Car类的3D检测精度分别为79.73%、89.16%,运算时间仅需80 ms。此外,基于机器人操作系统实现了算法的在线检测,验证了算法的工程实用性。
Objective The low detection accuracy of the perception system in autonomous vehicles will seriously affect the reliability of autonomous vehicle and the safety of passengers.The traditional LiDAR-based three-dimensional(3D)object detection algorithms,such as the rule-based clustering method highly relies on hand-designed features probably be sub-optimal.Following the great advantages in deep learning for image field,a large body of literature to explore the application of this technology for 3DLiDAR point clouds.Among them,point-based methods directly use raw point clouds as the input of the detection model,and the further point sampling(FPS)algorithm is applied to sample a set of keypoints from raw point clouds,keypoints groups neighbor raw points to extract the feature for object detection.However,the proportion of foreground points(points in 3Dbounding box)in keypoints collected through FPS algorithm are relatively low,especially for the remote object,foreground points almost totally lost in FPS(Fig.1).Foreground points contain the important 3Dspace location information of objects,a low proportion of foreground points in keypoints will hurt the detection accuracy.To this end,we propose a semantic segmentation based twostage 3Dobject detection algorithm named Seg-RCNN(segmentation based region-convolution neural networks),in which we propose a novel further point sampling strategy(segFPS)for sampling keypoints,and a segmentation network(SegNet)for semantic segmentation of foreground points and background points(Fig.4).Methods Seg-RCNN is a two-stage 3Dobject detector(Fig.2),in the first stage,the raw points are first voxelized as voxel-wise features,then the sparse 3DCNN is adopted to extract voxel features,the output of sparse3DCNN squeeze into 2DCNN for further feature propagation,and then box proposals are generated in 2Dbirds eye view feature map through anchor-based strategy.The SegNet output the foreground and background points segmentation results of point clouds.In the second stage,the SegFPS is adop
作者
赵亮
胡杰
刘汉
安永鹏
熊宗权
王宇
Zhao Liang;Hu Jie;Liu Han;An Yongpeng;Xiong Zongquan;Wang Yu(School of Automotive Engineering,Wuhan University of Technology,Wuhan,Hebei 430070,China;Hubei Key Laboratory of Advanced Technology for Automotive Components,Wuhan University of Technology,Wuhan,Hebei 430070,China;Hubei Collaborative Innovation Center for Automotive Components Technology,Wuhan University of Technology,Wuhan,Hebei 430070,China;Hubei Research Center for New Energy&Intelligent Connected Vehicle,Wuhan University of Technology,Wuhan,Hebei 430070,China)
出处
《中国激光》
EI
CAS
CSCD
北大核心
2021年第17期171-183,共13页
Chinese Journal of Lasers
基金
湖北省技术创新专项(2019AEA169)
湖北省科技重大专项(2020AAA001)。
关键词
遥感
自动驾驶
激光雷达
三维目标检测
深度学习
remote sensing
autonomous vehicle
LiDAR
three-dimensional object detection
deep learning
