[{"data":1,"prerenderedAt":189},["ShallowReactive",2],{"method-bosse_zlot2009_ctscan":3},{"method":4,"reference":52,"equipment":71,"figures":98,"results":99},{"id":5,"label":6,"shortName":7,"title":8,"year":9,"era":10,"cluster":11,"scope":12,"keyIdeaZh":13,"keyIdeaEn":14,"fulltextStatus":15,"publicationStatus":16,"recommendation":17,"constructionRelevance":18,"validationEnvironment":19,"strengths":22,"limitations":28,"sensors":35,"platform":37,"estimator":39,"association":40,"timeModel":41,"deskew":42,"loopClosure":43,"globalOptimization":44,"mapRepresentation":45,"prior":46,"outputGeometry":47,"compute":48,"codeUrl":49,"codeLicense":50,"relatedVersions":51},"bosse_zlot2009_ctscan","Bosse & Zlot, 2009","Continuous 3D scan-matching (Bosse and Zlot)","Continuous 3D scan-matching with a spinning 2D laser",2009,"classic","C04","odometry","本文處理移動中以旋轉 2D 雷射取得三維點雲時的運動畸變：每半圈（sweep）約需 1 秒，車輛在期間移動會使點雲局部變形。作者不停車，也不依賴里程計或 IMU，而是把 ICP 改成「掃描對掃描」的連續時間配準：先把點分入 0.5 至 8 m 的多解析度體素金字塔，以每個體素的一、二階矩求出橢球及其平面度與圓柱度，再在結合位置與形狀的 9 維描述空間中找對應；第二步不是求單一剛體轉換，而是每 0.2 秒取樣一個軌跡修正量，以匹配、平滑（加速度）與初始條件約束組成線性系統，配合 Cauchy 型穩健權重反覆求解；匹配約束在相鄰樣本間線性內插，最後以三次樣條重建連續軌跡。方法本身沒有迴圈閉合或全域最佳化，屬開迴路里程計。作者在工業園區與輕度林地以滑移轉向裝載機測試，報告局部精確的點雲與六自由度軌跡，但 MATLAB 實作約比即時慢五倍。","Sweep-to-sweep continuous-time ICP for a spinning 2D lidar on a moving skid-steer loader, without odometric or inertial sensors: multi-resolution voxel ellipsoids with plane and cylinder shape parameters are matched in a 9D position-shape space, and trajectory corrections sampled every 0.2 s are solved under match, smoothness and initial-condition constraints (linear interpolation between samples, cubic-spline reconstruction), giving an open-loop 6-DoF trajectory and unwarped point cloud; no loop closure, about five times slower than real time in MATLAB.","full_text_reviewed","peer_reviewed_published","background","實驗平台是營建常見的 Bobcat S185 滑移轉向裝載機，但測試場域為平坦鋪面的工業園區（有護柱與貨櫃）與輕度林地越野路段（Sec. III），不是營建工地、隧道或建物。工業園區的參考軌跡來自另一台水平 2D 雷射的 2D SLAM（假設無俯仰、側傾與垂直位移），只能檢查平面運動；越野路段則以未說明方法的閉迴路軌跡作參考，所以本文沒有點雲幾何精度的獨立驗證。同一團隊後續的 Zebedee（zebedee2012）與 Northparkes 礦坑建圖（zlot_bosse2014_mine）才延伸到手持、室內與地下環境；後者明言其非剛性配準的原始版本出自本文。若要說本文是 GeoSLAM ZEB 或 Hovermap 等商用設備的技術來源，本次仍未找到可核證的原始依據，不宜寫入正文。",[20,21],"cross_site","independent_reference",[23,24,25,26,27],"Industrial compound (flat, paved; about 200 m trajectory): relative to a horizontal-laser 2D SLAM trajectory, translational error growth generally below 5% of distance travelled and rotational error below 0.3 deg per metre for 5-30 m windows (Sec. III, Fig. 5, Fig. 7a-b)","Off-road lightly wooded terrain (548 m, 8 min): translational errors generally below 1% of distance travelled with a rotational bias of about 0.1 deg per metre, relative to a closed-loop trajectory from an undescribed method (Sec. III, Fig. 7c-d, Fig. 9)","Reported operation at vehicle speeds up to 6 m\u002Fs and turn rates of 20 deg\u002Fs (Sec. III)","Single inexpensive 2D laser, no additional odometric sensors (Sec. I)","Undistorted bollards and containers visible in a 40 m map segment (Sec. III, Fig. 6)",[29,30,31,32,33,34],"Open loop: no loop closure or global optimization in the method; the Fig. 6 map is explicitly without loop closure (Sec. III); consistent with elasticlidarfusion2018 Sec. II","Fast spot turns are the hardest motions; yaw is momentarily unobservable when the laser senses only the ground (Sec. III)","ICP convergence unreliable at high rotational velocities; sparse environments increase errors (Sec. III-IV)","Not real time: about 5 s per 1 s of data in MATLAB (Sec. IV); LOAM (Sec. II) describes Bosse and Zlot's methods ([3], [6], [22]) as requiring batch processing, and Wildcat (Sec. II) calls [22] an offline system","Weak references: the industrial reference is a 2D laser SLAM trajectory assuming zero pitch, roll and vertical translation; the off-road reference is a closed-loop trajectory from an undescribed method (Sec. III); (inference) it is probably derived from the same spinning-laser data and therefore not independent","Only two outdoor sites; no geometric accuracy assessment of the point cloud itself (Sec. III)",[36],"encoder on the spinning mount, described as accurate, giving each 2D scan's pose relative to the vehicle",[38],"vehicle: Bobcat S185 skid-steer loader, spinning laser above the cab (Sec. II, Fig. 1)","ICP variant for sweep-to-sweep matching: small trajectory corrections sampled every 0.2 s (six samples per sweep) solved from a stacked linear system of match, smoothness (acceleration) and initial-condition constraints after first-order linearization; robust M-estimator with Lorentzian\u002FCauchy weights; outer loop (correspondences) usually converges within about five iterations, inner loop (reweighted solve) drops from about seven to one or two iterations (Sec. II-B, Eq. 6-16)","voxel (not point) correspondences: points binned into a pyramid of 3D grids (0.5 m to 8 m cells, several offsets), each voxel timestamped with the mean time of its points; per-voxel first and second moments define an ellipsoid with cylinder-likeness c (Eq. 3) and plane-likeness p (Eq. 4); nearest neighbours searched in a 9D descriptor [alpha*mu; p*v1; c*v3] (Eq. 5) with alpha set to ten times the grid resolution and eigenvector signs fixed (v1 towards the sensor, v3 towards +z); planar matches constrain centroid offset along the normal and cylindrical matches perpendicular to the axis (Eq. 10-12), scaled by inverse square-root eigenvalues; normal-angle constraints derived but omitted in the implementation (Sec. II-A, II-B)","continuous-time trajectory T(tau) per sweep; corrections estimated at samples every 0.2 s, match constraints linearly interpolated between the two nearest samples (Eq. 13), and a cubic spline used to reconstitute a smooth continuous trajectory (Sec. II-B). This resolves the citing-paper discrepancy: elasticlidarfusion2018 describes the linear interpolation of constraints, wildcat2022 groups the paper with spline methods.","central aim: the recovered sweep trajectory unwarps motion-distorted sweeps; points measured much later than the earliest point in a voxel are excluded to limit distortion (Sec. II-A, Fig. 4)","none in the method; the authors mention appearance-based loop-closure investigations with preliminary results (Fig. 10) and place globally consistent mapping outside the paper's scope (Sec. I, III, IV)","none in the method; the off-road experiment compares the open-loop trajectory with a globally consistent closed-loop trajectory from a method 'not described in this paper' (Sec. III)","multi-resolution voxel ellipsoids (moments) for matching; output is the concatenated unwarped point cloud (Sec. II-A, Fig. 2, Fig. 6)","none from external sensors; prior trajectories from a motion model (sweep a: previous sweep motion; sweep b: deceleration to zero at sweep end); IMU or odometry seeding possible but not used (Sec. II-B)","open-loop 6-DoF sensor trajectory and locally consistent unwarped 3D point cloud; no covariance or map-accuracy metric reported (Sec. III-IV)","about 5 s of computation per 1 s of data in MATLAB on a 3.2 GHz Pentium 4 (about five times slower than real time); authors foresee real time after porting to C++ (Sec. IV)",null,"not_verified",[],{"id":5,"kind":53,"shortName":7,"title":8,"authors":54,"year":9,"venue":57,"venueType":58,"publisher":59,"volumeIssuePages":60,"doi":61,"arxivId":49,"url":62,"firstPublicDate":63,"publicationStatus":16,"metadataStatus":64,"fulltextStatus":15,"era":10,"classicReason":65,"codeUrl":49,"cluster":11,"topics":66,"mdpi":67,"verification":68,"label":6,"fulltextRoute":69,"versionRead":70,"addedByCensus":67},"method",[55,56],"Michael Bosse","Robert Zlot","2009 IEEE International Conference on Robotics and Automation (ICRA)","conference","IEEE","4312-4319","10.1109\u002Frobot.2009.5152851","https:\u002F\u002Fieeexplore.ieee.org\u002Fdocument\u002F5152851","2009-05","metadata_verified","principle reused: origin of the CSIRO continuous-time sweep-matching line. Zlot and Bosse 2014 (Sec. 4.1, reference 'Bosse and Zlot, 2009b') state that the original version of their non-rigid registration algorithm was introduced in this paper; Elastic LiDAR Fusion adopts its multi-resolution ellipsoid surfels (elasticlidarfusion2018 Sec. III-IV, ref. [9]); Wildcat reuses its plane-likeness score ([22, Eq. 4], wildcat2022 Sec. IV-A) and describes its odometry as an online implementation of concepts from this paper and Zebedee (Sec. II); LOAM names Bosse and Zlot ([3], [6], [22]) as the approach closest to its own (loam2014 Sec. II).",[11],false,"corrected","NTU institutional (Chrome)","version of record, IEEE Xplore full-text HTML (ICRA 2009, pp. 4312-4319); reference list not displayed on IEEE Xplore; Algorithm 1 pseudo-code body not rendered in the HTML (only its title), its content is described in the prose",[72,78,83,87,92],{"category":73,"model":74,"canonical":74,"role":75,"dataset":49,"specs":76,"locator":77},"lidar","SICK LMS291","method input","commercial 2D laser range finder spun about its centre scan line at 0.5 Hz; 75 Hz scan rate, 1 deg angular resolution; 150 scans per revolution; a half-revolution 'sweep' has 13,500 points; hemispherical field of view","Sec. II; Fig. 1",{"category":79,"model":80,"canonical":80,"role":75,"dataset":49,"specs":81,"locator":82},"other","encoder on the spinning mount","described as accurate; gives laser scan poses relative to the vehicle","Sec. II",{"category":84,"model":85,"canonical":85,"role":75,"dataset":49,"specs":86,"locator":77},"platform","Bobcat S185 skid-steer loader","spinning laser mounted above the cab, spin axis facing forward; skid-steer, can turn in place",{"category":73,"model":88,"canonical":88,"role":89,"dataset":49,"specs":90,"locator":91},"stationary 2D laser scanner mounted horizontally on the vehicle's bucket (model not stated)","reference or ground truth","processed by the authors' 2D SLAM framework [1] assuming zero pitch, roll and vertical translation","Sec. III",{"category":93,"model":94,"canonical":94,"role":95,"dataset":49,"specs":96,"locator":97},"compute","3.2 GHz Pentium 4 processor (MATLAB)","compute for runtime","about 5 s of processing per 1 s of data","Sec. IV",[],{"totalRows":100,"groupCount":101,"groups":102,"others":188},5,2,[103,160],{"slug":104,"group":105,"sourceId":5,"sourceLabel":6,"table":106,"selfRows":107,"metrics":108,"seqs":121,"entrants":130,"cells":134,"outcomes":145,"locators":150,"hardware":154,"wordings":155,"notes":156},"bosse-zlot2009-ctscan-text-sec-iii","bosse_zlot2009_ctscan:Text Sec. III","Text Sec. III",4,[109,114,117,119],{"label":110,"unit":111,"statistic":112,"alignment":113},"growth of translational errors, generally expected to be less than this share of distance travelled","%","not_reported","first-pose",{"label":115,"unit":116,"statistic":112,"alignment":113},"rotational errors per metre travelled","deg\u002Fm",{"label":118,"unit":111,"statistic":112,"alignment":113},"translational errors as share of distance travelled",{"label":120,"unit":116,"statistic":112,"alignment":113},"rotational bias",[122,126],{"dataset":123,"sequence":124,"environment":125},"authors' spinning-laser data (industrial)","about 200 m traverse","outdoor industrial compound (paved)",{"dataset":127,"sequence":128,"environment":129},"authors' spinning-laser data (off-road)","548 m traverse","outdoor off-road, lightly wooded",[131],{"name":132,"methodId":5,"linkable":133,"proposed":133,"self":133},"3D sweep-matching (continuous-time ICP)",true,[135,138,141,142],[136,136,136,100,136,136,137,137,136],0,-1,[136,139,136,140,139,136,137,137,139],1,0.3,[136,101,139,139,101,139,137,137,101],[136,143,139,144,143,101,137,137,101],3,0.1,[146,147,148,149],"upper bound stated in text ('less than 5%')","upper bound stated in text ('less than 0.3 deg per meter')","upper bound stated in text ('generally less than 1%')","approximate ('about 0.1 deg\u002Fm')",[151,152,153],"Sec. III; Fig. 7a-b","Sec. III; Fig. 7c","Sec. III; Fig. 7d",[],[],[157,158,159],"industrial compound, flat paved, about 200 m; sliding windows of 5 to 30 m aligned at window start; reference = 2D SLAM trajectory of a separate horizontal laser (assumes zero pitch, roll and vertical translation); values summarise box plots in Fig. 7a-b","industrial compound, flat paved, about 200 m; sliding windows of 5 to 30 m aligned at window start; reference = 2D SLAM trajectory of a separate horizontal laser; values summarise box plots in Fig. 7a-b","off-road lightly wooded terrain, 548 m, 8 min; reference = globally consistent closed-loop trajectory from a method not described in the paper; values summarise box plots in Fig. 7c-d",{"slug":161,"group":162,"sourceId":5,"sourceLabel":6,"table":163,"selfRows":139,"metrics":164,"seqs":170,"entrants":175,"cells":178,"outcomes":180,"locators":182,"hardware":183,"wordings":185,"notes":186},"bosse-zlot2009-ctscan-text-sec-iv","bosse_zlot2009_ctscan:Text Sec. IV","Text Sec. IV",[165],{"label":166,"unit":167,"statistic":168,"alignment":169},"processing time per second of data","s per 1 s of data","mean","not_applicable",[171],{"dataset":172,"sequence":173,"environment":174},"authors' spinning-laser data","all","outdoor",[176],{"name":177,"methodId":5,"linkable":133,"proposed":133,"self":133},"3D sweep-matching (continuous-time ICP), MATLAB implementation",[179],[136,136,136,100,136,136,136,137,136],[181],"approximate ('around 5 seconds')",[97],[184],"MATLAB on a 3.2 GHz Pentium 4 processor",[],[187],"average processing cost of the ICP sweep-matching in MATLAB",[],1790510655049]