PSM SLAM (Probabilistic Surfel Map)
PSM SLAM 以機率面元地圖(Probabilistic Surfel Map)結合逐影格與對模型兩類 RGB-D 視覺 SLAM。地圖中每個點帶有三維位置與 3×3 共變異、強度與其變異量以及法向,新觀測以兩個高斯分布相乘的方式融合,正確關聯的點其不確定度會快速下降,不可靠的點則被移除,因此地圖只保留稀疏而可靠的點。前端沿用 σ-DVO 的光度與幾何混合權重,把每一影格對齊到由全域地圖可見點與新關鍵影格觀測組成的 Keyframe PSM;後端以 g2o 交替最佳化關鍵影格之間的位姿約束,以及以不確定度加權的位姿對地圖點約束(光度式光束法平差)。需要稠密網格時,再以地圖點為控制點調整各關鍵影格的深度圖後融合。
本頁內容
Keyframe dense RGB-D SLAM with a Probabilistic Surfel Map: sparse points carrying position covariance and intensity variance are fused by Gaussian products; frames are aligned to a keyframe PSM with sigma-DVO weighting, and the back end alternates pose-pose graph optimization with uncertainty-weighted pose-point (photometric BA) constraints; dense meshes are produced on demand by deforming keyframe depth maps toward the PSM.
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 | RGB-D camera (TUM RGB-D real sequences; ICL-NUIM synthetic sequences with noise; sensor models not named) |
|---|---|
| 原文測試平台 | real RGB-D sequences (TUM RGB-D; capture platform not described in the paper)、simulation (ICL-NUIM living room with noise) |
| 狀態估計 | Keyframe-based dense visual odometry with the sigma-DVO hybrid weighting (Student-t for photometric, sensor-noise model for geometric residuals), aligning each frame to a Keyframe PSM; back end in g2o that alternates pose-pose graph optimization with uncertainty-weighted pose-point constraints (photometric bundle adjustment on PSM points), usually converging within 10 iterations |
| 資料關聯 | PSM points projected into the keyframe define active points; the Keyframe PSM merges them with back-projected new observations (pruned by position uncertainty and sampled at 20%); frame residuals are photometric and depth differences at projected points |
| 時間表示 | discrete poses (keyframes and frames attached to them) |
| 去畸變 | 不適用 (RGB-D input) |
| 迴圈閉合 | Nearest-neighbour search of keyframe poses in a kd-tree (as in sigma-DVO) adds pose-pose constraints |
| 全域最佳化 | g2o graph optimization alternating pose-pose and pose-point (PSM) constraints at the end of SLAM, with outlier-edge removal |
| 地圖表示 | Probabilistic Surfel Map: sparse points with 3D position and 3x3 covariance, intensity and intensity variance, and normal; updated by multiplying Gaussian distributions so that uncertainties of consistently associated points shrink; unreliable points pruned |
| 先驗資訊 | none |
| 可輸出幾何 | sparse globally consistent PSM; dense point cloud or mesh on request by deforming each keyframe depth map toward nearby PSM points (Gaussian-weighted KNN) and fusing |
| 計算需求 | CPU, multi-threaded, built on the open-source C++ DVO SLAM; workstation with Intel Xeon E5 @ 2.4 GHz; 20% point sample rate gives 30 to 40% less computation time than DVO and sigma-DVO on the same hardware (Sec. 6, 6.2) |
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 |
|---|---|---|---|---|---|
| 運算硬體 | workstation with Intel Xeon E5 @ 2.4GHz | 執行運算平台 | 未標示 | multi-threaded CPU implementation built on the open-source DVO SLAM; GPU acceleration is mentioned only as a possible improvement (Sec. 6) | (Yan et al., 2017, Sec. 6) |
作者報告的優勢與限制
優勢
- Visual odometry on 11 TUM sequences: average ATE 0.088 m and RPE 0.034 m/s versus 0.161 m and 0.042 m/s for sigma-DVO (Table 1)
- Full SLAM ATE lowest among the compared systems on fr1/desk (0.016 m), fr1/360 (0.055 m) and fr1/room (0.051 m) (Table 2)
- ICL-NUIM living room: lowest average ATE (0.024 m versus 0.034 m for ElasticFusion) and average surface error equal to ElasticFusion (0.012 m) (Tables 4 and 5)
- Sparse sampling (20%) cuts computation time by 30 to 40% compared with DVO and sigma-DVO on the same hardware (Sec. 6.2)
限制
- Can be confused and follow a wrong motion when a large part of the scene moves independently of the camera, for example a hallway with a moving crowd (Sec. 7)
- The sparse PSM leaves blank pixels; dense output requires a separate keyframe deformation and fusion step with higher error than the PSM alone (Sec. 6.3; Fig. 6)
- Less accurate than sigma-DVO SLAM on fr3/office (0.031 m versus 0.015 m) (Tables 2 and 3)
- The pose-point map optimization is run at the end of SLAM rather than continuously (Sec. 3 system overview; Sec. 5.2)
營建工程相關證據
論文未在施工現場測試,評估限於 TUM RGB-D 辦公室序列與 ICL-NUIM 合成客廳,目標應用為擴增實境。以不確定度加權的稀疏地圖在 CPU 上降低計算量,並可於需要時輸出稠密點雲,對室內擴增實境檢核有參考價值;但作者指出大量移動物體會使方法失效,且本次查核未找到公開程式碼(推論)。
原文驗證環境:公開基準、模擬
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 5 個比較組,合計 42 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 1 組列在最後,並連到性能比較頁。
Yan et al., 2017 · Table 1 本方法 22 筆
表格設定(擷取紀錄原文):TUM RGB-D visual odometry only (no back-end); RMSE computed with the TUM benchmark scripts; PSM VO RPE on fr1/desk2 printed as 0.50 (likely 0.050) (Yan et al., 2017, Table 1)
ATE [m],TUM RGB-D · fr1/360
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Yan et al., 2017 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yan et al., 2017, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| sigma-DVO (visual odometry) | 0.229 m | (Yan et al., 2017, Table 1) |
| PSM VO (visual odometry)本方法原文提出 | 0.113 m | (Yan et al., 2017, Table 1) |
Yan et al., 2017 · Table 3 本方法 8 筆
指標ATE [m]
表格設定(擷取紀錄原文):Complete SLAM ATE of sigma-DVO SLAM and PSM SLAM on TUM RGB-D (keyframe counts in the same table not extracted) (Yan et al., 2017, Table 3)
ATE [m],TUM RGB-D · fr1/xyz
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Yan et al., 2017 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yan et al., 2017, Table 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| sigma-DVO SLAM | 0.016 m | (Yan et al., 2017, Table 3) |
| PSM SLAM本方法原文提出 | 0.011 m | (Yan et al., 2017, Table 3) |
Yan et al., 2017 · Table 2 本方法 4 筆
指標ATE [m]
表格設定(擷取紀錄原文):Complete SLAM absolute trajectory error on TUM RGB-D; '-' = no result given (Yan et al., 2017, Table 2)
ATE [m],TUM RGB-D · fr1/desk
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Yan et al., 2017 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yan et al., 2017, Table 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| RGB-D SLAM [4] (Endres et al.) | 0.023 m | (Yan et al., 2017, Table 2) |
| Kintinuous [33] | 0.037 m | (Yan et al., 2017, Table 2) |
| MRSMap [28] | 0.043 m | (Yan et al., 2017, Table 2) |
| ElasticFusion [34] | 0.02 m | (Yan et al., 2017, Table 2) |
| DVO SLAM [14] | 0.021 m | (Yan et al., 2017, Table 2) |
| sigma-DVO SLAM | 0.019 m | (Yan et al., 2017, Table 2) |
| PSM SLAM本方法原文提出 | 0.016 m | (Yan et al., 2017, Table 2) |
Yan et al., 2017 · Table 4 本方法 4 筆
指標mean distance from points to nearest ground-truth surface (m)
表格設定(擷取紀錄原文):Surface reconstruction accuracy on ICL-NUIM living room with noise: mean distance from reconstructed points to the nearest ground-truth surface (m) (Yan et al., 2017, Table 4)
mean distance from points to nearest ground-truth surface (m),ICL-NUIM · lr kt0
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Yan et al., 2017 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yan et al., 2017, Table 4)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| RGB-D SLAM [4] (Endres et al.) | 0.044 m | (Yan et al., 2017, Table 4) |
| Kintinuous [33] | 0.011 m | (Yan et al., 2017, Table 4) |
| MRSMap [28] | 0.061 m | (Yan et al., 2017, Table 4) |
| DVO SLAM [14] | 0.032 m | (Yan et al., 2017, Table 4) |
| ElasticFusion [34] | 0.007 m | (Yan et al., 2017, Table 4) |
| PSM SLAM本方法原文提出 | 0.006 m | (Yan et al., 2017, Table 4) |
其他比較組
列出其餘 1 個比較組
來源
Yan et al., 2017
(2017)Dense Visual SLAM with Probabilistic Surfel MapIEEE Transactions on Visualization and Computer Graphics (ISMAR 2017 special issue), 23(11):2389-2398
同儕審查已出版已讀全文近十年查證後修正