來(lái)源丨h(huán)ttps://zhuanlan.zhihu.com/p/508859024過(guò)去很多年激光雷達(dá)的車規(guī)標(biāo)準(zhǔn)和高昂價(jià)格是阻礙其量產(chǎn)落地的主要因素,最近兩三年隨著速騰、禾賽、大疆、圖達(dá)通、Luminar等廠家混合固態(tài)激光雷達(dá)的量產(chǎn),新勢(shì)力車企、互聯(lián)網(wǎng)車企陸續(xù)發(fā)布與交付了基于激光雷達(dá)的車型,比如:小鵬P5、蔚來(lái)ET7/ET5、集度概念車、威馬M7、智己L7、高合HiPhiZ、沙龍機(jī)甲龍、極狐HBT,混合固態(tài)激光雷達(dá)即將進(jìn)入批量量產(chǎn)的前夜。后續(xù)隨著各大廠商智能電動(dòng)車型的大規(guī)模量產(chǎn)與交付,混合固態(tài)激光雷達(dá)可能將會(huì)是主流車型的標(biāo)配。LiDAR感知、定位、建圖、預(yù)測(cè)算法功能的開(kāi)發(fā)將在車企/供應(yīng)商ADAS團(tuán)隊(duì)中占比越來(lái)越多,不再僅僅是一個(gè)輔助/真值系統(tǒng)的存在。最近疫情在家,對(duì)過(guò)去幾年學(xué)習(xí)、積累的LiDAR目標(biāo)檢測(cè)算法(不包含傳統(tǒng)算法、車道線、FreeSpace檢測(cè))論文做了總結(jié),共計(jì)有54篇論文及代碼,有些是基礎(chǔ)網(wǎng)絡(luò)算法,有些經(jīng)典的、最新的算法也可作為工程落地的參考方案。基于激光雷達(dá)點(diǎn)云的3D目標(biāo)檢測(cè)算法有很多種方法:傳統(tǒng)聚類方法,點(diǎn)云、體素化、柱狀化,RangeView、BirdEyeView,多幀、多視圖,OneStage、TwoStage,AnchorBased、AnchroFree、關(guān)鍵點(diǎn)、中心點(diǎn)、Voting、與分割結(jié)合、結(jié)合反射強(qiáng)度與線束角、轉(zhuǎn)為深度圖,知識(shí)蒸餾、Transformer、Atteintion、半監(jiān)督,2DCNN、3D稀疏卷積、圖卷積,與Camera圖像數(shù)據(jù)數(shù)據(jù)融合、特征融合。從現(xiàn)階段角度,激光雷達(dá)本身還有很多工程問(wèn)題(布置、噪聲、標(biāo)定、同步、畸變、補(bǔ)償、安全)需要嘗試和解決,還有一個(gè)難點(diǎn)是網(wǎng)絡(luò)模型在嵌入式平臺(tái)的部署與優(yōu)化。但是對(duì)于目標(biāo)檢測(cè)算法本身,還是先基于CNN、BEV、AnchorBased/中心點(diǎn)為基礎(chǔ)算法完成工程落地,后續(xù)逐漸升級(jí)到以Transformer/Fusion框架的大感知框架。先以LiDAR/Camera后融合為主,可能的話,逐漸走向前融合的方案。題目:3DSSD: Point-based 3D Single Stage Object Detector名稱:3DSSD:基于點(diǎn)的 3D 單級(jí)物體檢測(cè)器論文:https://arxiv.org/abs/2002.10187代碼:https://github.com/tomztyang/3DSSD題目:AFDet: Anchor Free One Stage 3D Object Detection名稱:AFDet:無(wú)錨的一級(jí) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2006.12671題目:Associate-3Ddet: Perceptual-to-Conceptual Association for 3D Point Cloud Object Detection名稱:Associate-3Ddet:3D 點(diǎn)云對(duì)象檢測(cè)的感知到概念關(guān)聯(lián)論文:https://arxiv.org/abs/2006.04356題目:Back to Reality: Weakly-supervised 3D Object Detection with Shape-guided Label Enhancement名稱:回到現(xiàn)實(shí):帶有形狀引導(dǎo)標(biāo)簽增強(qiáng)的弱監(jiān)督 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2203.05238代碼:https://github.com/wyf-ACCEPT/BackToReality題目:BEVDetNet: Bird's Eye View LiDAR Point Cloud based Real-time 3D Object Detection for Autonomous Driving名稱:BEVDetNet:基于鳥(niǎo)瞰 LiDAR 點(diǎn)云的自動(dòng)駕駛實(shí)時(shí) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2104.10780題目:BirdNet: a 3D Object Detection Framework from LiDAR information名稱:BirdNet:來(lái)自 LiDAR 信息的 3D 對(duì)象檢測(cè)框架論文:https://arxiv.org/abs/1805.01195題目:BirdNet+: End-to-End 3D Object Detection in LiDAR Bird's Eye View名稱:BirdNet+:LiDAR 鳥(niǎo)瞰圖中的端到端 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2003.04188題目:Canonical Voting: Towards Robust Oriented Bounding Box Detection in 3D Scenes名稱:規(guī)范投票:在 3D 場(chǎng)景中實(shí)現(xiàn)穩(wěn)健的定向邊界框檢測(cè)論文:https://arxiv.org/abs/2011.12001代碼:https://github.com/qq456cvb/CanonicalVoting題目:CenterNet3D: An Anchor Free Object Detector for Point Cloud名稱:用于自動(dòng)駕駛的無(wú)錨物體檢測(cè)器論文:https://arxiv.org/abs/2007.07214代碼:https://github.com/wangguojun2018/CenterNet3d題目:Center-based 3D Object Detection and Tracking名稱:基于中心的3D目標(biāo)檢測(cè)和跟蹤論文:https://arxiv.org/abs/2006.11275代碼:https://github.com/tianweiy/CenterPoint題目:CG-SSD: Corner Guided Single Stage 3D Object Detection from LiDAR Point Cloud名稱:CG-SSD:來(lái)自 LiDAR 點(diǎn)云的角引導(dǎo)單級(jí) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2202.11868題目:CIA-SSD: Confident IoU-Aware Single-Stage Object Detector From Point Cloud名稱:CIA-SSD:來(lái)自點(diǎn)云的自信的 IoU 感知單級(jí)目標(biāo)檢測(cè)器論文:https://arxiv.org/abs/2012.03015代碼:https://github.com/Vegeta2020/CIA-SSD題目:Class-balanced Grouping and Sampling for Point Cloud 3D Object Detection名稱:用于點(diǎn)云 3D 對(duì)象檢測(cè)的類平衡分組和采樣論文:https://arxiv.org/abs/1908.09492題目:Complex-YOLO: Real-time 3D Object Detection on Point Clouds名稱:Complex-YOLO:點(diǎn)云上的實(shí)時(shí) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/1803.06199代碼:https://github.com/AI-liu/Complex-YOLO題目:Improving 3D Object Detection with Channel-wise Transformer名稱:使用 Channel-wise Transformer 改進(jìn) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2108.10723題目:Deformable PV-RCNN: Improving 3D Object Detection with Learned Deformations名稱:可變形 PV-RCNN:通過(guò)學(xué)習(xí)變形改進(jìn) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2008.08766題目:End-to-End Pseudo-LiDAR for Image-Based 3D Object Detection名稱:用于基于圖像的 3D 對(duì)象檢測(cè)的端到端偽激光雷達(dá)論文:https://arxiv.org/abs/2004.03080代碼:https://github.com/mileyan/pseudo-LiDAR_e2e論文:https://arxiv.org/abs/1908.02990題目:FVNet: 3D Front-View Proposal Generation for Real-Time Object Detection from Point Clouds名稱:FVNet:用于從點(diǎn)云進(jìn)行實(shí)時(shí)對(duì)象檢測(cè)的 3D 前視圖建議生成論文:https://arxiv.org/abs/1903.10750題目:From Multi-View to Hollow-3D: Hallucinated Hollow-3D R-CNN for 3D Object Detection名稱:從多視圖到 Hollow-3D:用于 3D 對(duì)象檢測(cè)的幻覺(jué) Hollow-3D R-CNN論文:https://arxiv.org/abs/2107.14391題目:Object as Hotspots: An Anchor-Free 3D Object Detection Approach via Firing of Hotspots名稱:對(duì)象即熱點(diǎn):通過(guò)觸發(fā)熱點(diǎn)的無(wú)錨 3D 對(duì)象檢測(cè)方法論文:https://arxiv.org/abs/1912.12791題目:HVPR: Hybrid Voxel-Point Representation for Single-stage 3D Object Detection名稱:HVPR:用于單級(jí) 3D 對(duì)象檢測(cè)的混合體素點(diǎn)表示論文:https://arxiv.org/abs/2104.00902題目:Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds名稱:并非所有點(diǎn)都是平等的:學(xué)習(xí)用于 3D LiDAR 點(diǎn)云的高效基于點(diǎn)的檢測(cè)器論文:https://arxiv.org/abs/2203.11139代碼:https://github.com/yifanzhang713/IA-SSD題目:LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving名稱:LaserNet:用于自動(dòng)駕駛的高效概率 3D 對(duì)象檢測(cè)器論文:https://arxiv.org/abs/1903.08701名稱:LiDAR R-CNN:一種高效且通用的 3D 物體檢測(cè)器論文:https://arxiv.org/abs/2103.15297代碼:https://github.com/tusimple/LiDAR_RCNN題目:MLCVNet: Multi-Level Context VoteNet for 3D Object Detection名稱:MLCVNet:用于三維目標(biāo)檢測(cè)的多級(jí)上下文VoteNet論文:https://openaccess.thecvf.com/content_CVPR_2020/papers/Xie_MLCVNet_Multi-Level_Context_VoteNet_for_3D_Object_Detection_CVPR_2020_paper.pdf代碼:https://github.com/NUAAXQ/MLCVNet題目:End-to-End Multi-View Fusion for 3D Object Detection in LiDAR Point Clouds名稱:用于 LiDAR 點(diǎn)云中 3D 對(duì)象檢測(cè)的端到端多視圖融合論文:https://arxiv.org/abs/1910.06528題目:From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network名稱:從點(diǎn)到部分:使用部分感知和部分聚合網(wǎng)絡(luò)從點(diǎn)云進(jìn)行 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/1907.03670代碼:https://github.com/sshaoshuai/PointCloudDet3D題目:PIXOR: Real-time 3D Object Detection from Point Clouds名稱:PIXOR:點(diǎn)云的實(shí)時(shí) 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/1902.06326題目:3D Object Detection with Pointformer名稱:3D Object Detection with Pointformer論文:https://arxiv.org/abs/2012.11409代碼:https://github.com/Vladimir2506/Pointformer題目:Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud名稱:Point-GNN:用于點(diǎn)云中 3D 對(duì)象檢測(cè)的圖神經(jīng)網(wǎng)絡(luò)論文:https://arxiv.org/abs/2003.01251代碼:https://github.com/WeijingShi/Point-GNN題目:PointPillars: Fast Encoders for Object Detection from Point Clouds名稱:PointPillars:點(diǎn)云目標(biāo)檢測(cè)的快速編碼器論文:https://arxiv.org/abs/1812.05784論文:https://openaccess.thecvf.com/content_CVPR_2019/papers/Lang_PointPillars_Fast_Encoders_for_Object_Detection_From_Point_Clouds_CVPR_2019_paper.pdf代碼:https://github.com/nutonomy/second.pytorch題目:PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud名稱:PointRCNN:來(lái)自點(diǎn)云的 3D 對(duì)象建議生成和檢測(cè)論文:https://arxiv.org/abs/1812.04244代碼:https://github.com/sshaoshuai/PointRCNN題目:Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving名稱:來(lái)自視覺(jué)深度估計(jì)的偽激光雷達(dá):彌合自動(dòng)駕駛 3D 對(duì)象檢測(cè)的差距論文:https://arxiv.org/abs/1812.07179代碼:https://github.com/mileyan/pseudo_lidar題目:PU-Net: Point Cloud Upsampling Network名稱:PU-Net:點(diǎn)云上采樣網(wǎng)絡(luò)論文:https://openaccess.thecvf.com/content_cvpr_2018/papers/Yu_PU-Net_Point_Cloud_CVPR_2018_paper.pdf代碼:https://github.com/yulequan/PU-Net題目:Point-Voxel CNN for Efficient 3D Deep Learning名稱:用于高效 3D 深度學(xué)習(xí)的點(diǎn)體素 CNN論文:https://arxiv.org/abs/1907.03739主頁(yè):https://pvcnn.mit.edu/項(xiàng)目:https://developer.nvidia.com/blog/point-voxel-cnn-3d/題目:PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection名稱:PV-RCNN:用于 3D 對(duì)象檢測(cè)的點(diǎn)體素特征集抽象論文:https://arxiv.org/abs/1912.13192代碼:https://github.com/open-mmlab/OpenPCDet題目:PV-RCNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection名稱:PV-RCNN++:用于 3D 對(duì)象檢測(cè)的具有局部向量表示的點(diǎn)體素特征集抽象論文:https://arxiv.org/abs/2102.00463代碼:https://github.com/open-mmlab/OpenPCDet題目:RangeDet:In Defense of Range View for LiDAR-based 3D Object Detection名稱:RangeDet:為基于 LiDAR 的 3D 對(duì)象檢測(cè)保護(hù)范圍視圖論文:https://arxiv.org/abs/2103.10039題目:SA-Det3D: Self-Attention Based Context-Aware 3D Object Detection名稱:SA-Det3D:基于自注意力的上下文感知 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2101.02672代碼:https://github.com/AutoVision-cloud/SA-Det3D題目:SASA: Semantics-Augmented Set Abstraction for Point-based 3D Object Detection名稱:SASA:基于點(diǎn)的 3D 對(duì)象檢測(cè)的語(yǔ)義增強(qiáng)集抽象論文:https://arxiv.org/pdf/2201.01976.pdf代碼:https://github.com/blakechen97/SASA題目:Structure Aware Single-stage 3D Object Detection from Point Cloud名稱:基于點(diǎn)云的結(jié)構(gòu)感知單級(jí)三維目標(biāo)檢測(cè)論文:https://www4.comp.polyu.edu.hk/~cslzhang/paper/SA-SSD.pdf代碼:https://github.com/skyhehe123/SA-SSD題目:SECOND: Sparsely Embedded Convolutional Detection論文:https://www.mdpi.com/1424-8220/18/10/3337題目:SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud名稱:SE-SSD:來(lái)自點(diǎn)云的自集成單級(jí)目標(biāo)檢測(cè)器論文:https://arxiv.org/abs/2104.09804代碼:https://github.com/Vegeta2020/SE-SSD題目:SIENet: Spatial Information Enhancement Network for 3D Object Detection from Point Cloud名稱:SIENet:用于從點(diǎn)云進(jìn)行 3D 對(duì)象檢測(cè)的空間信息增強(qiáng)網(wǎng)絡(luò)論文:https://arxiv.org/abs/2103.15396題目:SS3D: Single Shot 3D Object Detector名稱:SS3D:?jiǎn)未?3D 物體檢測(cè)器論文:https://arxiv.org/abs/2004.14674題目:Embracing Single Stride 3D Object Detector with Sparse Transformer名稱:使用 Sparse Transformer 擁抱單步 3D 對(duì)象檢測(cè)器論文:https://arxiv.org/abs/2112.06375代碼:https://github.com/TuSimple/SST題目:STD: Sparse-to-Dense 3D Object Detector for Point Cloud名稱:STD:點(diǎn)云的稀疏到密集 3D 對(duì)象檢測(cè)器論文:https://arxiv.org/abs/1907.10471題目:TANet: Robust 3D Object Detection from Point Clouds with Triple Attention名稱:TANet:來(lái)自具有三重注意力的點(diǎn)云的穩(wěn)健 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/1912.05163題目:Deep Hough Voting for 3D Object Detection in Point Clouds名稱:用于點(diǎn)云中 3D 對(duì)象檢測(cè)的深度霍夫投票論文:https://arxiv.org/abs/1904.09664代碼:https://github.com/facebookresearch/votenet題目:Voxel Transformer for 3D Object Detection名稱:用于 3D 對(duì)象檢測(cè)的體素轉(zhuǎn)換器論文:https://arxiv.org/abs/2109.02497代碼:https://github.com/PointsCoder/VOTR題目:Voxel-FPN: multi-scale voxel feature aggregation in 3D object detection from point clouds名稱:Voxel-FPN:點(diǎn)云 3D 對(duì)象檢測(cè)中的多尺度體素特征聚合論文:https://arxiv.org/abs/1907.05286題目:VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection名稱:VoxelNet:基于點(diǎn)云的 3D 對(duì)象檢測(cè)的端到端學(xué)習(xí)論文:https://arxiv.org/abs/1711.06396代碼:https://github.com/qianguih/voxelnet題目:Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection名稱:Voxel R-CNN:邁向高性能基于體素的 3D 對(duì)象檢測(cè)論文:https://arxiv.org/abs/2012.15712
本文僅做學(xué)術(shù)分享,如有侵權(quán),請(qǐng)聯(lián)系刪文。