作者丨eyesighting@知乎来源丨https://zhuanlan.zhihu.com/p/508859024编辑丨3D视觉工坊前言 过去很多年激光雷达的车规标准和高昂价格是阻碍其量产落地的主要因素,最近两三年随着速腾、禾赛、大疆、图达通、Luminar等厂家混合固态激光雷达的量产,新势力车企、互联网车企陆续发布与交付了基于激光雷达的车型,比如:小鹏P5、蔚来ET7/ET5、集度概念车、威马M7、智己L7、高合HiPhiZ、沙龙机甲龙、极狐HBT,混合固态激光雷达即将进入批量量产的前夜。后续随着各大厂商智能电动车型的大规模量产与交付,混合固态激光雷达可能将会是主流车型的标配。LiDAR感知、定位、建图、预测算法功能的开发将在车企/供应商ADAS团队中占比越来越多,不再仅仅是一个辅助/真值系统的存在。最近疫情在家,对过去几年学习、积累的LiDAR目标检测算法(不包含传统算法、车道线、FreeSpace检测)论文做了总结,共计有54篇论文及代码,有些是基础网络算法,有些经典的、最新的算法也可作为工程落地的参考方案。基于激光雷达点云的3D目标检测算法有很多种方法:传统聚类方法,点云、体素化、柱状化,RangeView、BirdEyeView,多帧、多视图,OneStage、TwoStage,AnchorBased、AnchroFree、关键点、中心点、Voting、与分割结合、结合反射强度与线束角、转为深度图,知识蒸馏、Transformer、Atteintion、半监督,2DCNN、3D稀疏卷积、图卷积,与Camera图像数据数据融合、特征融合。从现阶段角度,激光雷达本身还有很多工程问题(布置、噪声、标定、同步、畸变、补偿、安全)需要尝试和解决,还有一个难点是网络模型在嵌入式平台的部署与优化。但是对于目标检测算法本身,还是先基于CNN、BEV、AnchorBased/中心点为基础算法完成工程落地,后续逐渐升级到以Transformer/Fusion框架的大感知框架。先以LiDAR/Camera后融合为主,可能的话,逐渐走向前融合的方案。算法论文3DSSD 题目:3DSSD: Point-based 3D Single Stage Object Detector名称:3DSSD:基于点的 3D 单级物体检测器论文:https://arxiv.org/abs/2002.10187代码:https://github.com/tomztyang/3DSSDAFDet 题目:AFDet: Anchor Free One Stage 3D Object Detection名称:AFDet:无锚的一级 3D 对象检测论文:https://arxiv.org/abs/2006.12671Associate-3DDet 题目:Associate-3Ddet: Perceptual-to-Conceptual Association for 3D Point Cloud Object Detection名称:Associate-3Ddet:3D 点云对象检测的感知到概念关联论文:https://arxiv.org/abs/2006.04356BackReality 题目:Back to Reality: Weakly-supervised 3D Object Detection with Shape-guided Label Enhancement名称:回到现实:带有形状引导标签增强的弱监督 3D 对象检测论文:https://arxiv.org/abs/2203.05238代码:https://github.com/wyf-ACCEPT/BackToRealityBEVDetNet 题目:BEVDetNet: Bird's Eye View LiDAR Point Cloud based Real-time 3D Object Detection for Autonomous Driving名称:BEVDetNet:基于鸟瞰 LiDAR 点云的自动驾驶实时 3D 对象检测论文:https://arxiv.org/abs/2104.10780BirdNet 题目:BirdNet: a 3D Object Detection Framework from LiDAR information名称:BirdNet:来自 LiDAR 信息的 3D 对象检测框架论文:https://arxiv.org/abs/1805.01195BirdNet+ 题目:BirdNet+: End-to-End 3D Object Detection in LiDAR Bird's Eye View名称:BirdNet+:LiDAR 鸟瞰图中的端到端 3D 对象检测论文:https://arxiv.org/abs/2003.04188CanonicalVoting题目:Canonical Voting: Towards Robust Oriented Bounding Box Detection in 3D Scenes名称:规范投票:在 3D 场景中实现稳健的定向边界框检测论文:https://arxiv.org/abs/2011.12001代码:https://github.com/qq456cvb/CanonicalVotingCenterNet3D 题目:CenterNet3D: An Anchor Free Object Detector for Point Cloud名称:用于自动驾驶的无锚物体检测器论文:https://arxiv.org/abs/2007.07214代码:https://github.com/wangguojun2018/CenterNet3dCenterPoint 题目:Center-based 3D Object Detection and Tracking名称:基于中心的3D目标检测和跟踪论文:https://arxiv.org/abs/2006.11275代码:https://github.com/tianweiy/CenterPointCG-SSD 题目:CG-SSD: Corner Guided Single Stage 3D Object Detection from LiDAR Point Cloud名称:CG-SSD:来自 LiDAR 点云的角引导单级 3D 对象检测论文:https://arxiv.org/abs/2202.11868CIA-SSD 题目:CIA-SSD: Confident IoU-Aware Single-Stage Object Detector From Point Cloud名称:CIA-SSD:来自点云的自信的 IoU 感知单级目标检测器论文:https://arxiv.org/abs/2012.03015代码:https://github.com/Vegeta2020/CIA-SSDClassBalanced-GS 题目:Class-balanced Grouping and Sampling for Point Cloud 3D Object Detection名称:用于点云 3D 对象检测的类平衡分组和采样论文:https://arxiv.org/abs/1908.09492Complex-YOLO 题目:Complex-YOLO: Real-time 3D Object Detection on Point Clouds名称:Complex-YOLO:点云上的实时 3D 对象检测论文:https://arxiv.org/abs/1803.06199代码:https://github.com/AI-liu/Complex-YOLOCT3D 题目:Improving 3D Object Detection with Channel-wise Transformer名称:使用 Channel-wise Transformer 改进 3D 对象检测论文:https://arxiv.org/abs/2108.10723Deformable-PV-RCNN 题目:Deformable PV-RCNN: Improving 3D Object Detection with Learned Deformations名称:可变形 PV-RCNN:通过学习变形改进 3D 对象检测论文:https://arxiv.org/abs/2008.08766E2E-PL 题目:End-to-End Pseudo-LiDAR for Image-Based 3D Object Detection名称:用于基于图像的 3D 对象检测的端到端伪激光雷达论文:https://arxiv.org/abs/2004.03080代码:https://github.com/mileyan/pseudo-LiDAR_e2eFast-Point-RCNN 题目:Fast Point R-CNN名称:快速点 R-CNN论文:https://arxiv.org/abs/1908.02990FVNet 题目:FVNet: 3D Front-View Proposal Generation for Real-Time Object Detection from Point Clouds名称:FVNet:用于从点云进行实时对象检测的 3D 前视图建议生成论文:https://arxiv.org/abs/1903.10750Hollow3D-RCNN 题目:From Multi-View to Hollow-3D: Hallucinated Hollow-3D R-CNN for 3D Object Detection名称:从多视图到 Hollow-3D:用于 3D 对象检测的幻觉 Hollow-3D R-CNN论文:https://arxiv.org/abs/2107.14391HotSpotNet 题目:Object as Hotspots: An Anchor-Free 3D Object Detection Approach via Firing of Hotspots名称:对象即热点:通过触发热点的无锚 3D 对象检测方法论文:https://arxiv.org/abs/1912.12791HVPR 题目:HVPR: Hybrid Voxel-Point Representation for Single-stage 3D Object Detection名称:HVPR:用于单级 3D 对象检测的混合体素点表示论文:https://arxiv.org/abs/2104.00902IS-SSD 题目:Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds名称:并非所有点都是平等的:学习用于 3D LiDAR 点云的高效基于点的检测器论文:https://arxiv.org/abs/2203.11139代码:https://github.com/yifanzhang713/IA-SSDLaserNet 题目:LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving名称:LaserNet:用于自动驾驶的高效概率 3D 对象检测器论文:https://arxiv.org/abs/1903.08701Lidar-RCNN 名称:LiDAR R-CNN:一种高效且通用的 3D 物体检测器论文:https://arxiv.org/abs/2103.15297代码:https://github.com/tusimple/LiDAR_RCNNMLCVNet 题目:MLCVNet: Multi-Level Context VoteNet for 3D Object Detection名称:MLCVNet:用于三维目标检测的多级上下文VoteNet论文:https://openaccess.thecvf.com/content_CVPR_2020/papers/Xie_MLCVNet_Multi-Level_Context_VoteNet_for_3D_Object_Detection_CVPR_2020_paper.pdf代码:https://github.com/NUAAXQ/MLCVNetMVF 题目:End-to-End Multi-View Fusion for 3D Object Detection in LiDAR Point Clouds名称:用于 LiDAR 点云中 3D 对象检测的端到端多视图融合论文:https://arxiv.org/abs/1910.06528PartA2Net 题目:From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network名称:从点到部分:使用部分感知和部分聚合网络从点云进行 3D 对象检测论文:https://arxiv.org/abs/1907.03670代码:https://github.com/sshaoshuai/PointCloudDet3DPIXOR 题目:PIXOR: Real-time 3D Object Detection from Point Clouds名称:PIXOR:点云的实时 3D 对象检测论文:https://arxiv.org/abs/1902.06326Pointformer题目:3D Object Detection with Pointformer名称:3D Object Detection with Pointformer论文:https://arxiv.org/abs/2012.11409代码:https://github.com/Vladimir2506/PointformerPoint-GNN 题目:Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud名称:Point-GNN:用于点云中 3D 对象检测的图神经网络论文:https://arxiv.org/abs/2003.01251代码:https://github.com/WeijingShi/Point-GNNPointPillars 题目:PointPillars: Fast Encoders for Object Detection from Point Clouds名称:PointPillars:点云目标检测的快速编码器论文:https://arxiv.org/abs/1812.05784论文:https://openaccess.thecvf.com/content_CVPR_2019/papers/Lang_PointPillars_Fast_Encoders_for_Object_Detection_From_Point_Clouds_CVPR_2019_paper.pdf代码:https://github.com/nutonomy/second.pytorchPointRCNN 题目:PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud名称:PointRCNN:来自点云的 3D 对象建议生成和检测论文:https://arxiv.org/abs/1812.04244代码:https://github.com/sshaoshuai/PointRCNNPseudo-LiDAR 题目:Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving名称:来自视觉深度估计的伪激光雷达:弥合自动驾驶 3D 对象检测的差距论文:https://arxiv.org/abs/1812.07179代码:https://github.com/mileyan/pseudo_lidarPU-Net题目:PU-Net: Point Cloud Upsampling Network名称:PU-Net:点云上采样网络论文:https://openaccess.thecvf.com/content_cvpr_2018/papers/Yu_PU-Net_Point_Cloud_CVPR_2018_paper.pdf代码:https://github.com/yulequan/PU-NetPoint-Voxel题目:Point-Voxel CNN for Efficient 3D Deep Learning名称:用于高效 3D 深度学习的点体素 CNN论文:https://arxiv.org/abs/1907.03739主页:https://pvcnn.mit.edu/项目:https://developer.nvidia.com/blog/point-voxel-cnn-3d/PV-RCNN 题目:PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection名称:PV-RCNN:用于 3D 对象检测的点体素特征集抽象论文:https://arxiv.org/abs/1912.13192代码:https://github.com/open-mmlab/OpenPCDetPV-RCNN++题目:PV-RCNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection名称:PV-RCNN++:用于 3D 对象检测的具有局部向量表示的点体素特征集抽象论文:https://arxiv.org/abs/2102.00463代码:https://github.com/open-mmlab/OpenPCDetRangeDet 题目:RangeDet:In Defense of Range View for LiDAR-based 3D Object Detection名称:RangeDet:为基于 LiDAR 的 3D 对象检测保护范围视图论文:https://arxiv.org/abs/2103.10039SA-Det3D 题目:SA-Det3D: Self-Attention Based Context-Aware 3D Object Detection名称:SA-Det3D:基于自注意力的上下文感知 3D 对象检测论文:https://arxiv.org/abs/2101.02672代码:https://github.com/AutoVision-cloud/SA-Det3DSASA 题目:SASA: Semantics-Augmented Set Abstraction for Point-based 3D Object Detection名称:SASA:基于点的 3D 对象检测的语义增强集抽象论文:https://arxiv.org/pdf/2201.01976.pdf代码:https://github.com/blakechen97/SASASA-SSD题目:Structure Aware Single-stage 3D Object Detection from Point Cloud名称:基于点云的结构感知单级三维目标检测论文:https://www4.comp.polyu.edu.hk/~cslzhang/paper/SA-SSD.pdf代码:https://github.com/skyhehe123/SA-SSDSECOND 题目:SECOND: Sparsely Embedded Convolutional Detection名称:第二:稀疏嵌入卷积检测论文:https://www.mdpi.com/1424-8220/18/10/3337SE-SSD题目:SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud名称:SE-SSD:来自点云的自集成单级目标检测器论文:https://arxiv.org/abs/2104.09804代码:https://github.com/Vegeta2020/SE-SSDSIENet 题目:SIENet: Spatial Information Enhancement Network for 3D Object Detection from Point Cloud名称:SIENet:用于从点云进行 3D 对象检测的空间信息增强网络论文:https://arxiv.org/abs/2103.15396SS3D 题目:SS3D: Single Shot 3D Object Detector名称:SS3D:单次 3D 物体检测器论文:https://arxiv.org/abs/2004.14674相关课程:国内首个3D缺陷检测教程:理论、源码与实战SST 题目:Embracing Single Stride 3D Object Detector with Sparse Transformer名称:使用 Sparse Transformer 拥抱单步 3D 对象检测器论文:https://arxiv.org/abs/2112.06375代码:https://github.com/TuSimple/SSTSTD 题目:STD: Sparse-to-Dense 3D Object Detector for Point Cloud名称:STD:点云的稀疏到密集 3D 对象检测器论文:https://arxiv.org/abs/1907.10471TANet 题目:TANet: Robust 3D Object Detection from Point Clouds with Triple Attention名称:TANet:来自具有三重注意力的点云的稳健 3D 对象检测论文:https://arxiv.org/abs/1912.05163VoteNet 题目:Deep Hough Voting for 3D Object Detection in Point Clouds名称:用于点云中 3D 对象检测的深度霍夫投票论文:https://arxiv.org/abs/1904.09664代码:https://github.com/facebookresearch/votenetVOTR题目:Voxel Transformer for 3D Object Detection名称:用于 3D 对象检测的体素转换器论文:https://arxiv.org/abs/2109.02497代码:https://github.com/PointsCoder/VOTRVoxel-FPN题目:Voxel-FPN: multi-scale voxel feature aggregation in 3D object detection from point clouds名称:Voxel-FPN:点云 3D 对象检测中的多尺度体素特征聚合论文:https://arxiv.org/abs/1907.05286VoxelNet题目:VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection名称:VoxelNet:基于点云的 3D 对象检测的端到端学习论文:https://arxiv.org/abs/1711.06396代码:https://github.com/qianguih/voxelnetVoxel-RCNN 题目:Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection名称:Voxel R-CNN:迈向高性能基于体素的 3D 对象检测论文:https://arxiv.org/abs/2012.15712