Long-term 3D map maintenance in dynamic environments
本文提出以單一 3D 雷射進行長期定位與建圖的系統，重點是地圖隨時間的維護。新點雲先以 libpointmatcher 的 ICP 配準到全域稀疏點雲地圖，系統再依可見性假設逐點更新地圖點為動態的貝氏機率：若新讀值出現在原地圖點後方（1 度圓錐內），代表雷射穿過了該點位置，該點較可能已移動；更新時同時考慮兩射線夾角、入射角與距離雜訊。被判為動態的點再以雙向非剛性 ICP 估計逐點速度，不需物件模型或分群；P(Dyn) 也可作為 ICP 權重，以減少動態點造成的定位漂移。
本頁內容
Long-term 3D laser mapping that updates a per-point Bayesian dynamic probability from visibility in a sparse ICP map, estimates per-point velocities with dual non-rigid ICP, and can use P(Dyn) to down-weight dynamic points in registration.
技術屬性
欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。
| 感測輸入 | 3D LiDAR (Velodyne HDL-32E)、wheel odometry (prior alignment for registration) |
|---|---|
| 原文測試平台 | UGV (ARTOR, based on the LandShark system) |
| 狀態估計 | ICP registration of each point cloud to the global map with libpointmatcher, wheel odometry as prior; P(Dyn) can weight points in ICP to discount dynamic points (Sec. II; Sec. V.A) |
| 資料關聯 | ICP nearest neighbours via libnabo k-d trees; for dynamic inference, map points are associated with each new reading inside a 1 deg cone in spherical coordinates (Sec. III) |
| 時間表示 | discrete poses (one per scan) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | 原文未報告 |
| 全域最佳化 | none reported |
| 地圖表示 | sparse global point cloud with surface normals, timestamps and per-point Bayesian probability of being dynamic; a new point is added only if its nearest map point is farther than 0.3 m; time history of per-point velocities kept (Sec. III-V) |
| 先驗資訊 | prior map from earlier sessions (the first survey is the exploration phase) |
| 可輸出幾何 | static-scene point map (P(Dyn) < 0.5), dynamic points with velocity vectors, and maps of dynamic-element occurrence, average speed and heading (Figs. 5, 10, 11) |
| 計算需求 | single laptop with a four-core Intel Core i7 and 4 GB RAM: registration 6 to 11 Hz, map maintenance about 2 Hz with maps up to 600,000 points, velocity estimation 0.03 s on average (Sec. V.D) |
使用設備
原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。
| 類別 | 型號（原文寫法） | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| LiDAR | Velodyne HDL-32E | 方法輸入 | 未標示 | roughly 70,000 points per 360 deg scan at 11 Hz; maximum range 80 m | (Pomerleau et al., 2014, Sec. V) |
| 載具平台 | ARTOR | 方法輸入 | 未標示 | maximum speed 3.5 m/s, typically 1 m/s in crowded environments; large sensor suite | (Pomerleau et al., 2014, Sec. V; Fig. 1) |
| 運算硬體 | four-core Intel Core i7 | 執行運算平台 | 未標示 | single laptop, 4 GB RAM | (Pomerleau et al., 2014, Sec. V.D) |
作者報告的優勢與限制
優勢
- Overall parking-lot segmentation error decreased from 20% to around 5% over nine surveys in three days against a night-time ground-truth map (Sec. V.A; Fig. 6).
- Weighting ICP by P(Dyn) solved localization drift caused by dynamic points accumulating in a 2 m layer above the ground (Sec. V.A).
- Median estimated speeds of 1.7 and 4.1 m/s for a walking and jogging pedestrian, and 1.9, 4.6 and 8.1 m/s for a minibus driven at about 2, 5.5 and 11 m/s; velocities estimated up to 22 m from the sensor (Sec. V.B).
- On a 1.3 km Zurich route surveyed three times over seven months, dynamic-element occurrence maps highlighted two construction sites, a felled tree and a busy intersection (Sec. V.C; Fig. 10).
限制
- Periodic dynamic objects such as trams and cars in fixed parking spaces make 'static' ambiguous; points that reach P(Dyn) >= 0.9 cannot return to static (Sec. III).
- Free space is modelled only where the laser previously returned a reading, not throughout the volume (Sec. I).
- Points near the ground, often lower parts of cars, were misclassified, so the static-point error rose over time (Sec. V.A).
- Velocity vectors are noisy at small displacements; the data rate was reduced to 8 Hz to handle 1.5 to 10 m/s (Sec. V.D).
- Trams were poorly detected because only their front or rear reveals motion (Sec. V.C).
- Minibus median speeds were lower than the targets because acceleration phases were included (Sec. V.B).
營建工程相關證據
未在工地內部驗證，但在蘇黎世 1.3 km、跨七個月的三次調查中，動態元素發生次數地圖標出兩處部分占用街道的施工區（Sec. V.C；Fig. 10）。逐點動態機率與可見性更新的做法，可用於重複掃描的工地中區分人員、機具等移動物體與新施作的靜態構件；但週期性出現的物體（如固定位置停放的機具）會使靜態的定義變得模糊（Sec. III）（推論）。
原文驗證環境：受控實驗
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 4 個比較組，合計 15 筆紀錄。
Pomerleau et al., 2014 · Text Sec. V.B 本方法 6 筆
表格設定（擷取紀錄原文）：Controlled street without traffic, robot parked, one moving object at a time; target speed not stated; for the minibus the acceleration phase is included, which lowers the median (Pomerleau et al., 2014, Text Sec. V.B)
median estimated speed，authors' ARTOR HDL-32E data (remote street) · pedestrian, Walking
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Pomerleau et al., 2014 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| dual non-rigid ICP velocity estimation本方法原文提出 | 1.7 m/s | (Pomerleau et al., 2014, Sec. V.B; Fig. 9) |
Koide et al., 2021a · Table I 本方法 4 筆
資料集與序列PASCO Mobile Measurement System outdoor dataset (released via SMRT-AIST) · outdoor sequence (about 20 min)
表格設定（擷取紀錄原文）：PASCO Mobile Measurement System outdoor dataset (about 200 m x 400 m, 20 min); ground truth is the 3D LiDAR position tracked by static total stations (translation only); errors given as mean +/- std; RTE uses a modified routine that aligns each sub-trajectory within the evaluation window; the proposed rows start from LOAM odometry with 5 m keyframes (Koide et al., 2021a, Table I)
ATE [m], mean +/- std = 5.248 +/- 5.313，PASCO Mobile Measurement System outdoor dataset (released via SMRT-AIST) · outdoor sequence (about 20 min)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Koide et al., 2021a 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ethzasl_icp_mapping [33]本方法 | 5.248 m | (Koide et al., 2021a, Table I) |
Pomerleau et al., 2014 · Text Sec. V.D 本方法 3 筆
資料集與序列authors' ARTOR data · all experiments
表格設定（擷取紀錄原文）：Module rates with input run at recorded rate; registration downsamples points and uses wheel odometry as prior (Pomerleau et al., 2014, Text Sec. V.D)
registration module rate，authors' ARTOR data · all experiments
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Pomerleau et al., 2014 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| registration module (libpointmatcher ICP)本方法原文提出硬體：single laptop, four-core Intel Core i7, 4 GB RAM | 無數值僅報告範圍註記（擷取紀錄）：range 6 to 11 Hz | (Pomerleau et al., 2014, Sec. V.D) |
Pomerleau et al., 2014 · Text Sec. V.A 本方法 2 筆
資料集與序列authors' ARTOR HDL-32E data (parking lot) · first survey
表格設定（擷取紀錄原文）：Hospital visitor parking lot, nine surveys over three days; classification compared with a night-time ground-truth map (static if a ground-truth point lies within 0.15 m) (Pomerleau et al., 2014, Text Sec. V.A)
overall classification error (start of the series)，authors' ARTOR HDL-32E data (parking lot) · first survey
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Pomerleau et al., 2014 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Bayesian dynamic-point classification本方法原文提出 | 20%有附註註記（擷取紀錄）：approximate (text: from 20%) | (Pomerleau et al., 2014, Sec. V.A; Fig. 6) |
來源
Pomerleau et al., 2014
（2014）Long-term 3D map maintenance in dynamic environments2014 IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, pp. 3712-3719
同儕審查已出版已讀全文經典查證後修正
相關版本
- repository copy：HAL hal-01143106 (submitted version) https://hal.science/hal-01143106
- attributed software：ethz-asl/ethzasl_icp_mapping; cited to this paper by interactive_slam, but the repository does not cite it https://github.com/ethz-asl/ethzasl_icp_mapping