View management for lifelong visual maps

问题描述

SLAM系统在运行是会不断创建新的视图用于全局定位和回环检测。如果视图数量不断增加，会破坏SLAM算法的实时性。为了满足终身定位的需要，本文提出了一种修剪视图的方法，来保持视图数量在限制数量以下。

去除原理

 Algorithm 1 Find views for deletion   Require: Set  V  of all the views   1: if  ∣ V ∣ ≤ M I N − VIEWS then   2: return   3:  V keep  ←  newlyCreatedViews()   4: for  v  in  ( V \ V keep  )  do   5:  score ⁡ [ v ] ← ( W 1 × r e l o c + W 2 × n − o b s − c u r m a x − o b s + W 3 × n − o b s − r u n s n − r u n s )  6: if score [ v ] >  SCORE_THRESHOLD then   7:  V keep  ← V keep  ∪ { v }  8:  V delete  ← V \ V keep   9:  sort  ( V delete  ,  score  )  # Sort by score   10:  D ← ∅  11: for  v  in  V delete   do   12:  if numNearestNeighbors  ( v , D ) <  NN.THRESHOLD   then   13:  V keep  ← V keep  ∪ { v }  14: else   15:  D ← D ∪ { v }  16:  V delete  ← V \ V keep   17: return  V delete  \begin{aligned} &\begin{array}{l} \text { Algorithm 1 Find views for deletion } \\ \hline \text { Require: Set } \mathscr{V} \text { of all the views } \\ \text { 1: if }|\mathscr{V}| \leq \mathrm{MIN}_{-} \text {VIEWS then } \\ \text { 2: return } \\ \text { 3: } \mathscr{V}_{\text {keep }} \leftarrow \text { newlyCreatedViews() } \\ \text { 4: for } v \text { in }\left(\mathscr{V} \backslash \mathscr{V}_{\text {keep }}\right) \text { do } \\ \text { 5: } \operatorname{score}[v] \leftarrow \left(W_{1} \times r e l o c+W_{2} \times \frac{n_{-} o b s_{-} c u r}{m a x_{-} o b s}+W_{3} \times \frac{n_{-} o b s_{-} r u n s}{n_{-} r u n s}\right) \\ \text { 6: if score}[ v]>\text { SCORE\_THRESHOLD then } \\ \text { 7: } \mathscr{V}_{\text {keep }} \leftarrow \mathscr{V}_{\text {keep }} \cup\{v\} \\ \text { 8: } \mathscr{V}_{\text {delete }} \leftarrow \mathscr{V} \backslash \mathscr{V}_{\text {keep }}\\ \text { 9: } \text { sort }\left(\mathscr{V}_{\text {delete }}, \text { score }\right) \text { \# Sort by score } \\ \text { 10: } D \leftarrow \varnothing \\ \text { 11: for } v \text { in } \mathscr{V}_{\text {delete }} \text { do } \\ \text { 12: } \text { if numNearestNeighbors }(v, D)<\text { NN.THRESHOLD } \\ \text { then } \\ \text { 13: } \mathscr{V}_{\text {keep }} \leftarrow \mathscr{V}_{\text {keep }} \cup\{v\} \\ \text { 14: else } \\ \text { 15: } D \leftarrow D \cup\{v\} \\ \text { 16: } \mathscr{V}_{\text {delete }} \leftarrow \mathscr{V} \backslash \mathscr{V}_{\text {keep }} \\ \text { 17: return } \mathscr{V}_{\text {delete }} \\ \hline \end{array} \end{aligned} ​ Algorithm 1 Find views for deletion  Require: Set V of all the views  1: if ∣V∣≤MIN−​VIEWS then  2: return  3: Vkeep ​← newlyCreatedViews()  4: for v in (V\Vkeep ​) do  5: score[v]←(W1​×reloc+W2​×max−​obsn−​obs−​cur​+W3​×n−​runsn−​obs−​runs​) 6: if score[v]> SCORE_THRESHOLD then  7: Vkeep ​←Vkeep ​∪{v} 8: Vdelete ​←V\Vkeep ​ 9:  sort (Vdelete ​, score ) # Sort by score  10: D←∅ 11: for v in Vdelete ​ do  12:  if numNearestNeighbors (v,D)< NN.THRESHOLD  then  13: Vkeep ​←Vkeep ​∪{v} 14: else  15: D←D∪{v} 16: Vdelete ​←V\Vkeep ​ 17: return Vdelete ​​​​
去除视图的代码如上，${\mathcal{V}}_{\text{keep }}$为应保留的视图子集， $\mathcal{V}$为所有视图的集合。 得分超过阈值的视图会加入 ${\mathcal{V}}_{\text{keep }}$中。得分计算中， $W_1$， $W_2$，$W_3$为三个部分的权重，如果视图用于重定位， $reloc$为1，否则为0。 $n_{-}runs$是总视图数， $n_{-}ob{s}_{-}runs$是能观测到当前视图的视图数。 $n_{-}ob{s}_{-}cur$是当前能观测到的视图数， $ma{x}_{-}obs$是最多观测到的视图数。 ${\mathcal{V}}_{\text{delete}}$中视图周围的视图数小于 $NN.THRESHOLD$，将该视图加回 ${\mathcal{V}}_{\text{keep }}$。