-
摘要: 提出了一种新的基于集中式处理结构的有约束多传感器控制算法.该算法将多目标均方误差界作为传感器控制的代价函数.为了应用信息不等式得到该误差界, 2阶最优子模式分配测度被用于度量状态集和其估计集间的误差, 并采用δ-广义标签多伯努利滤波器执行多目标Bayes递推.混合罚函数法和复合形法被用来降低求解该有约束优化问题的计算量.仿真结果表明对于由多个不同观测性能传感器组成的带约束条件的控制系统, 本方法的跟踪精度显著优于柯西-施瓦茨散度法; 并且当传感器个数较多时, 混合罚函数和复合形法的计算时间相比穷尽搜索法显著缩短而跟踪精度损失很小.Abstract: The paper proposes a new constrained multi-sensor control algorithm based on the centralized processing architecture. In this method, a multi-target mean-square error bound is served as cost function of sensor control. In order to derive the bound by using the information inequality, the error between state set and its estimation is measured by the 2nd-order optimal sub-pattern assignment metric while the multi-target Bayes recursion is performed by using a δ-generalized labeled multi-Bernoulli filter. Mixed penalty function method and complex method are used to reduce the computation cost of solving the constrained optimization problem. Simulation results show that for the constrained multi-sensor control system with different observation performance, our method significantly outperforms the Cauchy-Schwarz divergence method in tracking precision. Besides, when the number of sensors is relatively large, the computation time of the mixed penalty function and complex methods is much shorter than that of the exhaustive search method at the expense of completely acceptable loss of tracking accuracy.
-
Key words:
- Multi-sensor control /
- labeled random finite set (RFS) /
- multi-target tracking /
- Bayesian estimation /
- error bounds
-
在目标跟踪[1-2]领域, 传感器控制是指在通信带宽、能量或计算量等约束条件下, 通过使传感器执行一定操作以提高目标检测和估计精度.由于受目标个数和状态不确定性、测量噪声、漏检、杂波、非线性以及实时性等影响, 传感器控制相比传统控制问题更为复杂, 因而受到了广泛关注[3-5].
近年来, 随着随机有限集(Random finite set, RFS)[6]和标签RFS[7-10]在多目标滤波中的广泛应用, 相应的传感器控制方法也获得了迅猛发展. RFS的最大优势是可以将多目标跟踪和传感器控制统一描述为一个Bayes框架下部分可观测的马尔科夫决策过程(Partially observed Markov decision process, POMDP)[11].对于多目标单传感器控制问题, 近些年已取得了一些成果[12-18]; 而对于多传感器场景, 由于其复杂性显著增加, 目前仅有最大化柯西-施瓦茨(Cauchy-Schwarz, CS)散度法[19]和最小化势和状态的后验期望误差(Posterior expected error of cardinality and states, PEECS)法[20].这两种方法均基于分布式融合结构, 故必须采用广义协方差交叉(Generalized covariance intersection, GCI)[21-22]准则获得融合后的多传感器后验密度.但在GCI中各个传感器权重的设定一直没有可靠的依据, 往往凭经验而定.例如, 当传感器观测性能相同或接近时, 可将各个权重设为相同; 反之, 当传感器观测性能相差较大时, 则很难找到设定各个权重的有效方法.此外, 对于多传感器控制问题, 若采用穷尽搜索法寻找最优控制命令, 其计算代价会随传感器个数的增加而显著增长.针对该问题, Wang等[20]建议采用坐标下降法[23]来降低无约束多传感器控制的计算代价.并且除文献[18]外, 均没有考虑传感器控制的约束条件.
针对上述问题, 通过采用δ-广义标签多伯努利(δ-generalized labeled multi-bernoulli, δ-GLMB)滤波器[7]执行多目标Bayes递推, 本文提出了一种以多目标均方误差(Mean-square error, MSE)界作为代价函数且带约束条件的多传感器控制算法.此处采用2阶最优子模式分配(Optimal sub-pattern assignment, OSPA)测度[24]定义该误差距离.为了获得该误差的下界, 需要将传统应用于矢量的信息不等式扩展到RFS观测[25].由于本文误差界以直到当前时刻的具体观测集为条件, 其包含了关于目标状态实现的有用信息, 因此相比仅仅依赖于运动和观测模型的无条件界[26], 它能够提供更加准确的多目标在线估计性能, 因此也更加适用于传感器实时控制.本文方法无需采用GCI准则, 建议采用混合罚函数法[27]和复合形法[28]来降低其计算代价.仿真结果表明对于观测性能差异较大且带约束条件的多传感器控制系统, 本方法相比基于GCI的CS散度法的多目标跟踪精度有明显提高.并且对于传感器较多的场景, 在保证多目标跟踪精度损失很小的条件下, 采用所建议的子优化算法相比采用穷尽搜索法计算时间会显著缩短.
1. 问题描述
为便于区分, 用斜体表示未加标签的量(例如将传统的状态矢量、测量矢量和其相应集合记为$ {x} $, $ z $, $ {X} $和$ {Z} $)而粗斜体表示带标签的量(例如将标签状态矢量和其集合记为$ {\pmb x} $和$ {\pmb X} $).令$ {\delta _Y}(X) $、$ {1_Y}(X) $和$ {p^{\pmb X}} $分别表示广义Kroneker、包含指示和多目标指数函数,
$$ \begin{align} {\delta _Y}(X) = \begin{cases} {1, }& X = Y\\ {0, }&\mbox{其他} \end{cases} \end{align} $$ (1) $$ \begin{align} {1_Y}(X) = \begin{cases} {1, }& X \subseteq Y \\ {0, }&\mbox{其他} \end{cases} \end{align} $$ (2) $$ \begin{align} {p^{\pmb X}} = \begin{cases} {\prod\limits_{{\pmb x} \in {\pmb X}} {p({\pmb x})} , }&{{\pmb X} \ne \varnothing } \\ {1, }&{{\pmb X} = \varnothing } \end{cases} \end{align} $$ (3) 通常将$ {1_Y}\left( {\{ x\} } \right) $简写为$ {1_Y}(x) $.
记$ \hat x $为由$ m $元素测量集$ {Z_m} $得到的$ x $的无偏估计, $ f\left( {x, {Z_m}} \right) $为$ \ x $和$ {Z_m} $的联合密度.假设正则条件成立且${\partial ^2}\log f\left( {x,{Z_m}} \right)/\partial {x^i}\partial {x^j}$存在, 可得基于RFS观测的信息不等式为[29]
$$ \begin{align} {\int_{{\mathbb{Z}_m}} {\int_{{\mathbb{X}_1}} {f\left( {x, {Z_m}} \right){{\left( {{x^l} - {{\hat x}^l}} \right)}^2}{\rm d}x{\rm d}{z_{1:m}}} } \ge {{\left[ {J_m^{ - 1}} \right]}^{l, l}}, }\\{l = 1, \cdots, L} \end{align} $$ (4) 式中, $ {z_{1:m}} = {z_1}, \cdots, {z_m} $, $ \ L $为$ x $的维数, $ {x^l} $和$ {\hat x^l} $为矢量$ x $和$ \hat x $的第$ l $个分量, $ {\mathbb{X}_1}\times {\mathbb{Z}_m} $为$ \left( {x, {Z_m}} \right) $所在空间, $ J_m $是以$ |Z| = m $为条件的$ L\times L $维Fisher信息矩阵(Fisher information matrix, FIM), 其中$ \text{ }|\text{ }\begin{matrix} \cdot \\ \end{matrix}\text{ }|\text{ } $表示集合的势,
$$ \begin{align} &{{\left[ {{J}_{m}} \right]}^{i, j}} = -{{E}_{f}} \left[ \frac{{{\partial }^{2}}\log f\left( x, {{Z}_{m}} \right)}{\partial {{x}^{i}}\partial {{x}^{j}}} \right] = \\ &\quad -\int_{{{\mathbb{Z}}_{m}}} \int_{{{\mathbb{X}}_{1}}}f\left( x, {{Z}_{m}} \right)\frac{{{\partial }^{2}}\log f\left( x, {{Z}_{m}} \right)}{\partial {{x}^{i}}\partial {{x}^{j}}}{\rm d}x{\rm d}{{z}_{1:m}}, \\&\hskip40mm {i, j = 1, \cdots, L} \end{align} $$ (5) 当且仅当$ f\left( x, {{Z}_{m}} \right) $服从指数分布族时, 式(4)中的等号成立[29].
为表述简洁, 本文将时间下标省略, 用下标`$ + $'和`$ - $'分别表示一步预测量和前一时刻量.
假设$ s $个独立可控传感器对多个独立运动的目标进行观测, 目标可能随时间消亡和新生, 传感器可能收到杂波和目标测量或发生漏检.将多目标状态集$ {\pmb X} $建模为空间$ \mathbb{X}\times \mathbb{L} $上的标签RFS, 其中$ \mathbb{X} $和$ \mathbb{L} $分别表示状态空间和离散标签空间.单目标状态$ {\pmb x} = \left( x, \ell \right)\in {\pmb X} $的存活概率和转移密度分别为$ {{p}_{S}}\left( x, \ell \right) $和$ f\left( \left. x, \ell \right|{{x}_{-}}, {{\ell }_{-}} \right){{\delta }_{{{\ell }_{-}}}}(\ell ) $.
将传感器$ i $($ i = 1, \cdots, s $)的测量集$ {{Z}^{i}} $建模为空间$ {{\mathbb{Z}}^{i}} $上的RFS, 记单测量矢量为$ {{z}^{i}}\in {{Z}^{i}} $, 控制矢量为$ {{u}^{i}}\in {{\mathbb{U}}^{i}} $, 其中, $ {{\mathbb{U}}^{i}} $为控制指令空间.杂波建模为泊松RFS, 其强度为
$$ \begin{align} {{\kappa }^{i}}\left( {{z}^{i}} \right) = {{\lambda }^{i}}f_{\kappa }^{i}\left( {{z}^{i}} \right) \end{align} $$ (6) 其中, $ {{\lambda }^{i}} $和$ f_{\kappa }^{i}\left( {{z}^{i}} \right) $分别为平均每帧杂波个数和单杂波密度.根据文献[6], 似然函数$ {{g}^{i}}\left( {{Z}^{i}}\left| {\pmb X}, {{u}^{i}} \right. \right) $可写为
$$ \begin{align} &{{g}^{i}}\left( {{Z}^{i}}\left| {\pmb X}, {{u}^{i}} \right. \right) = \\&\qquad{{\rm e}^{-{{\lambda }^{i}}}}{{\left[ {{\kappa }^{i}} \right]}^{{{Z}^{i}}}}\sum\limits_{{{\theta }^{i}}\in {{\Theta }^{i}}}{{{\delta }_{{{({{\theta }^{i}})}^{-1}}(\{0:|{{Z}^{i}}|\})}}\left( \mathcal{L}({\pmb X}) \right)}\times\\&\qquad{{\left[ \psi _{{{Z}^{i}}, {{u}^{i}}}^{i}\left( \cdot ;{{\theta }^{i}} \right) \right]}^{{\pmb X}}} \end{align} $$ (7) 其中, $ \mathcal{L}({\pmb X}) $表示$ {\pmb X} $的标签集, $ {{\theta }^{i}} $表示关联映射$ \mathbb{L}\to \left\{ 0:\left| {{Z}^{i}} \right| \right\} = \left\{ 0, 1, \cdots, \left| {{Z}^{i}} \right| \right\} $, $ {{\theta }^{i}} $满足若$ {{\theta }^{i}}(j) = {{\theta }^{i}}({j}')>0 $则$ j = {j}' $, $ {{\Theta }^{i}} $为$ {{\theta }^{i}} $所在空间,
$$ \begin{align} &\psi _{{{Z}^{i}}, {{u}^{i}}}^{i}\left( x, \ell ;{{\theta }^{i}} \right) = {{\delta }_{0}}\left( {{\theta }^{i}}(\ell ) \right)\left( 1-p_{D}^{i}\left( x, \ell , {{u}^{i}} \right) \right)+\\& \quad \left( 1-{{\delta }_{0}}\left( {{\theta }^{i}}(\ell ) \right) \right)\frac{p_{D}^{i}\left( x, \ell , {{u}^{i}} \right){{g}^{i}}\left( \left. z_{{{\theta }^{i}}(\ell )}^{i} \right|x, \ell , {{u}^{i}} \right)}{{{\kappa }^{i}}\left( z_{{{\theta }^{i}}(\ell )}^{i} \right)} \end{align} $$ (8) 其中, $ p_{D}^{i}\left( x, \ell , {{u}^{i}} \right) $和$ {{g}^{i}}\left( \left. {{z}^{i}} \right|x, \ell , {{u}^{i}} \right) $分别为单目标检测概率和似然函数.
省略直到上一时刻的传感器累积测量集和控制命令序列.记$ {{Z}^{1:s}} = {{Z}^{1}}, \cdots, {{Z}^{s}} $和$ {{u}^{1:s}} = {{u}^{1}}, \cdots, {{u}^{s}} $.在Bayes框架下, 多传感器状态估计集$ \hat{{\pmb X}} $由多目标后验密度$ \pi \left( \left. {\pmb X} \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) $获得, 其递推公式为
$$ \begin{align} {{\pi }_+}({\pmb X}) = \int{f\left( \left. {\pmb X} \right|{{{\pmb X}}_{-}} \right)\pi \left( {{{\pmb X}}_{-}} \right)\delta {{{\pmb X}}_{-}}} \end{align} $$ (9) $$ \begin{align} \pi \left( \left. {\pmb X} \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) = \frac{\prod\nolimits_{i = 1}^{s}{{{g}^{i}}\left( \left. {{Z}^{i}} \right|{\pmb X}, {{u}^{i}} \right)}{{\pi }_+}({\pmb X})}{\int{\prod\nolimits_{i = 1}^{s}{{{g}^{i}}\left( \left. {{Z}^{i}} \right|{\pmb X}, {{u}^{i}} \right)}{{\pi }_+}({\pmb X})\delta {\pmb X}}} \end{align} $$ (10) 式中, $ {{\pi }_+}({\pmb X}) $和$ f\left( \left. {\pmb X} \right|{{{\pmb X}}_{-}} \right) $分别表示多目标先验密度和转移密度; $ \int{\cdot \delta {\pmb X}} $表示标签RFS积分, 对于函数$ b({\pmb X}) $其定义为:
$$ \begin{align} \int{b({\pmb X})\delta {\pmb X}} = \sum\limits_{n = 0}^{\infty }{\frac{1}{n!}\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\int_{{{\mathbb{X}}_{n}}}{b\left( {{{\pmb X}}_{n}} \right){\rm d}{{x}_{1:n}}}}} \end{align} $$ (11) 式中, $ {{x}_{1:n}} = {{x}_{1}}, \cdots, {{x}_{n}} $和$ {{\ell }_{1:n}} = {{\ell }_{1}}, \cdots, {{\ell }_{n}} $, $ {{\mathbb{X}}_{n}} $和$ {{\mathbb{L}}_{n}} $分别表示$ {{X}_{n}} $和$ {{\ell }_{1:n}} $所在空间.
对于Bayes跟踪系统, 多传感器控制的目的是在一定约束下选择合适的控制命令$ {{u}^{1:s}} $使得目标函数达到最优, 且$ {{u}^{1:s}} $的最优值需要在得到测量集$ {{Z}^{1:s}} $之前确定.故最优控制命令$ {{\left[ {{u}^{1:s}} \right]}^{*}} $可由下式获得:
$$ \begin{align} & {{\left[ {{u}^{1:s}} \right]}^{*}} = \underset{{{u}^{1:s}}\in {{\mathbb{U}}^{1:s}}}{\mathop{{\arg \min }/{\max }\;}}\, \text{ }\vartheta \left( {{u}^{1:s}};{{\pi }_+} \right) \\ & \text{s .t.}\; \; \begin{cases} {{\gamma }_{i}}\left( {{u}^{1:s}};{{\pi }_+} \right)\ge 0, & i = 1, \cdots, l \\ {{\nu }_{j}}\left( {{u}^{1:s}};{{\pi }_+} \right) = 0, & j = 1, \cdots, m \end{cases} \end{align} $$ (12) 式中, $ \vartheta \left( {{u}^{1:s}};{{\pi }_+} \right) $、$ {{\gamma }_{i}}\left( {{u}^{1:s}};{{\pi }_+} \right)\ge0 $ ($ i = 1, \cdots, l $)和$ {{\nu }_{j}}\left( {{u}^{1:s}};{{\pi }_+} \right) = 0 $ ($ j = 1, \cdots, m $)分别表示以多目标先验密度$ {{\pi }_+}({\pmb X}) $为条件的关于控制命令$ {{u}^{1:s}} $的目标函数、$ l $个不等式约束条件和$ m $个等式约束条件, $ {{u}^{1:s}}\in {{\mathbb{U}}^{1:s}} $表示$ {{u}^{1}}\in {{\mathbb{U}}^{1}}, \cdots, {{u}^{s}}\in {{\mathbb{U}}^{s}} $.称
$$ \begin{align} \begin{subarray}{l} \mathbb{U}_{D}^{1:s} = \left\{{{u}^{1:s}}\in {{\mathbb{U}}^{1:s}} \left|\begin{matrix} {{\gamma }_{i}}\left( {{u}^{1:s}};{{\pi }_+} \right)\ge 0, \text{ }i = 1, \cdots, l; \\ {{\nu }_{j}}\left( {{u}^{1:s}};{{\pi }_+} \right) = 0, \text{ }j = 1, \cdots, m . \\ \end{matrix}\right.\text{ }\right\} \end{subarray} \end{align} $$ (13) 为$ {{u}^{1:s}} $的可行域(允许集), $ \mathbb{U}_{D}^{1:s}\subseteq {{\mathbb{U}}^{1:s}} $.
2. 多传感器多目标MSE界
标签RFS框架下由$ s $个独立可控传感器获得的多目标Bayes估计的MSE可定义为
$$ \begin{array}{*{35}{l}} & {{\sigma }^{2}}\left( {{u}^{1:s}} \right)=E\left[ {{e}^{2}}\left( \boldsymbol{X},\widehat{\boldsymbol{X}} \right) \right]= \\ {} & \begin{align} & \ \ \ \ \int_{{{\mathbb{Z}}^{s}}}{\cdots }\int_{{{\mathbb{Z}}^{1}}}{\int_{\mathbb{X}\times \mathbb{L}}{f\left( \left. \boldsymbol{X},{{Z}^{1:s}} \right|{{u}^{1:s}} \right)}}{{e}^{2}}\left( \boldsymbol{X},\widehat{\boldsymbol{X}} \right)\delta \boldsymbol{X}\delta {{Z}^{1:s}}= \\ & \ \ \ \ \int_{{{\mathbb{Z}}^{s}}}{\cdots }\int_{{{\mathbb{Z}}^{1}}}{\int_{\mathbb{X}\times \mathbb{L}}{\prod\limits_{i=1}^{s}{{{g}^{i}}\left( {{Z}^{i}}|\boldsymbol{X},{{u}^{i}} \right)}}}\pi \left( \boldsymbol{X} \right){{e}^{2}}\left( \boldsymbol{X},\widehat{\boldsymbol{X}} \right)\delta \boldsymbol{X}\delta {{Z}^{1:s}} \\ \end{align} \\ \end{array} $$ (14) 式中, $ f\left( \left. {\pmb X}, {{Z}^{1:s}} \right|{{u}^{1:s}} \right) $表示以$ {{u}^{1:s}} $为条件的$ \left( {\pmb X}, {{Z}^{1:s}} \right) $的联合密度, $ e\left( {\pmb X}, \hat{{\pmb X}} \right) $表示$ {\pmb X} $和$ \hat{{\pmb X}} $的误差.
本文通过最小化$ {{\sigma }^{2}}\left( {{u}^{1:s}} \right) $的下界$ {{\underline{\sigma }}^{2}}\left( {{u}^{1:s}} \right) $来寻找多传感器最优控制量.为方便起见, $ {{\pi }_+}({\pmb X}) $作为默认条件而不再显式出现.最终式(12)可重写为
$$ \begin{align} & {{\left[ {{u}^{1:s}} \right]}^{*}} = \underset{{{u}^{1:s}} \in {{\mathbb{U}}^{1:s}}}{\mathop{{\arg \min }/{\max }}}\, \text{ }{{{\underline{\sigma }}}^{2}} \left( {{u}^{1:s}} \right) \\ & \text{s .t.}\; \; \begin{cases} {{\gamma }_{i}}\left( {{u}^{1:s}} \right)\ge 0, & i = 1, \cdots, l \\ {{\nu }_{j}}\left( {{u}^{1:s}} \right) = 0, & j = 1, \cdots, m \end{cases} \end{align} $$ (15) 为了得到$ {{\underline{\sigma }}^{2}}\left( {{u}^{1:s}} \right) $, 需要给定如下3点假设:
假设1. 误差$ e\left( {\pmb X}, \hat{{\pmb X}} \right) $由2阶OSPA测度定义[24], 见式(16).
$$ e\left( {{\bf{X}},{\bf{\hat X}}} \right) = \left\{ \begin{array}{l} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;0\;\;|\widehat {\bf{X}}|\; = \;|{\bf{X}}|\;\; = \;0\;\\ {\left( {\frac{{\tau \in \Gamma \mathop {\min }\limits_{\max (|\widehat {\bf{X}}|,|{\bf{X}}|)} \sum\limits_{i = 1}^{\min (|\widehat {\bf{X}}|,|{\bf{X}}|)} {\min \left( {{c^2},{{\left\| {{x_i} - {{\hat x}_{{\tau _i}}}} \right\|}^2}} \right)} + {c^2}{\rm{abs}}\left( {|\widehat {\bf{X}}| - |{\bf{X}}|} \right)}}{{\max \left( {|\widehat {\bf{X}}|,|{\bf{X}}|} \right)}}} \right)^{\frac{1}{2}}}\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;|\widehat {\bf{X}}| + |{\bf{X}}|\; > 0 \end{array} \right.$$ (16) 式(16)中$ {{\Gamma }_{n}} $表示由$ \{1, \cdots, n\} $的全排列构成的集合, $ \tau = \left\{{{\tau }_{1:n}}\right\} $表示$ {{\Gamma }_{n}} $中的元素, $ c $表示截断误差, $ \max (\cdot ) $、$ \min (\cdot ) $和\text{abs}$ (\cdot ) $表示取最大值, 最小值和绝对值, $ ||\cdot || $表示2范数.
假设2. 采用$ \delta $-GLMB滤波器执行多目标Bayes递推, 故多目标先验密度$ {{\pi }_+}({\pmb X}) $可写为
$$ \begin{align} {{\pi }_+}({\pmb X}) = \Delta ({\pmb X})\sum\limits_{(I, \xi )\in \mathcal{F}(\mathbb{L})\times \Xi }{{{\delta }_{I}}\left( \mathcal{L}({\pmb X}) \right)\omega _+^{(I, \xi )}{{\left[ p_+^{(\xi )} \right]}^{{\pmb X}}}} \end{align} $$ (17) 式(17)中$ \Delta ({\pmb X}) = {{\delta }_{|{\pmb X}|}}\left( \left| \mathcal{L}({\pmb X}) \right| \right) $表示标签区别指示器, $ I\in \mathcal{F}(\mathbb{L}) $表示航迹标签集, 其中$ \mathcal{F}(\mathbb{L}) $表示由空间$ \mathbb{L} $内的有限个子集构成的集合, $ \xi \in \Xi $表示历史关联映射, 权重$ \omega _+^{(I, \xi )} $表示$ (I, \xi ) $的预测概率, $ p_+^{(\xi )}\left( \cdot , \ell \right) $表示在映射$ \xi $下航迹$ \ell $的状态预测密度.
假设3. 最大后验(Maximum a posterior, MAP)检测和无偏估计准则, 即先根据MAP检测准则获得目标个数估计, 再获得该个数下每个目标的无偏状态估计.
引理1. 给定假设2和3, MAP检测准则判定$ |\hat{{\pmb X}}| = \hat{n} $ ($ \hat{n} = 0, 1, \cdots, \infty $)当且仅当$ {{Z}^{1:s}}\subseteq \mathbb{Z}_{{\hat{n}}}^{1:s} $, 如式(18), 其中, $ {{\mathbb{Z}}^{1:s}} = {{\mathbb{Z}}^{1}}\times \cdots \times {{\mathbb{Z}}^{s}} $表示$ s $个传感器的测量空间, $ \mathbb{Z}_{{\hat{n}}}^{1:s} = \mathbb{Z}_{{\hat{n}}}^{1}\times \cdots \times \mathbb{Z}_{{\hat{n}}}^{s} $表示将目标个数估计为$ \hat{n} $时所对应的$ s $个传感器的测量子空间, $ \mathbb{Z}_{0}^{1:s}, \mathbb{Z}_{1}^{1:s}, \cdots, \mathbb{Z}_{\infty }^{1:s} $为$ {{\mathbb{Z}}^{1:s}} $的划分, $ {{\theta }^{1:s}}\in {{\Theta }^{1:s}} $表示$ {{\theta }^{1}}\in {{\Theta }^{1}}, \cdots, {{\theta }^{s}}\in {{\Theta }^{s}} $, $ {{\mathcal{F}}_{n}}(\mathbb{L}) $表示由空间$ \mathbb{L} $内的有限个$ n $元素子集构成的集合, 如式(19).
引理1的证明见附录A.
记$ Z_{{{m}^{1:s}}}^{1:s} = Z_{{{m}^{1}}}^{1}, \cdots, Z_{{{m}^{s}}}^{s} $和$ \mathbb{Z}_{{{m}^{1:s}}}^{1:s} = \mathbb{Z}_{{{m}^{1}}}^{1}\times \cdots \times \mathbb{Z}_{{{m}^{s}}}^{s} $为$ Z_{{{m}^{1:s}}}^{1:s} $所在空间.
令$ q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $表示以$ {{u}^{1:s}} $为条件的空间$ {{\left( \mathbb{X}\times \mathbb{L} \right)}_{n}}\times \mathbb{Z}_{{{m}^{1:s}}}^{1:s} $上的联合密度, 根据Bayes公式可得:
$$ \begin{align} &q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) = \\&\qquad \frac{1}{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}\prod\limits_{i = 1}^{s}{{{g}^{i}}\left( \left. Z_{{{m}^{i}}}^{i} \right|{{{\pmb X}}_{n}}, {{u}^{i}} \right)}{{\pi }_+}\left( {{{\pmb X}}_{n}} \right) \end{align} $$ (20) 其中, $ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $为归一化因子, 见式(21), 式(21)中, $ z_{1:{{m}^{i}}}^{i} = z_{1}^{i}, \cdots, z_{{{m}^{i}}}^{i} $.
令$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}}\right) $表示$ q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $在空间$ {{\left( \mathbb{X}\times \mathbb{L} \right)}_{n}}\times \mathbb{Z}_{\hat{n}, {{m}^{1:s}}}^{1:s} $上的积分, 根据式(20)可得式(22).式(22)中积分区域中$ s $个传感器的测量子空间$ \mathbb{Z}_{\hat{n}, {{m}^{1:s}}}^{1:s} = \mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}\times \cdots \times \mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s} $可根据引理1获得.
将假设2和式(7)代入式(21)和(22)并根据文献[7]的引理12, $ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $最终可写为式(23)和(24).
$$ \begin{align} \mathbb{Z}_{{\hat{n}}}^{1:s} = \left\{ {{Z}^{1:s}}\subseteq {{\mathbb{Z}}^{1:s}}: \hat{n} = \underset{n} {\mathop{\arg \max }}\, \left( \sum\limits_{(I, \xi )\in {{\mathcal{F}}_{n}}(\mathbb{L})\times \Xi } {\sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}} {\omega _{{{Z}^{1:s}}, {{u}^{1:s}}}^{(\xi , {{\theta }^{1:s}})}(I)}} \right) \right\} \end{align} $$ (18) $$ \begin{align} \omega _{{{Z}^{1:s}}, {{u}^{1:s}}}^{(\xi , {{\theta }^{1:s}})}(I) = \frac{\omega _+^{(I, \xi )}{{\left\langle p_+^{(\xi )} \left( \cdot , \ell \right), \prod\limits_{i = 1}^{s} {{{\delta }_{{{({{\theta }^{i}})}^{-1}}(\{0:|{{Z}^{i}}|\})}} (I)\psi _{{{Z}^{i}}, {{u}^{i}}}^{i}\left( \cdot , \ell ; {{\theta }^{i}} \right)} \right\rangle }^{I}}}{\sum\limits_{(I, \xi ) \in \mathcal{F}(\mathbb{L})\times \mathbb{Z}} {\omega _+^{(I, \xi )}\sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}}{{{\left\langle \!p_+^{(\xi )} \left( \cdot , \ell \right), \prod\limits_{i = 1}^{s} {{{\delta }_{{{({{\theta }^{i}})}^{-1}}(\{0:|{{Z}^{i}}|\})}} (I)\psi _{{{Z}^{i}}, {{u}^{i}}}^{i}\left( \cdot , \ell ; {{\theta }^{i}} \right)} \right\rangle }^{I}}}}} \end{align} $$ (19) $$ \begin{align} &{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) = \sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\int_{\mathbb{Z}_{{{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{{{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{n}}}}}}{\prod\limits_{i = 1}^{s}{{{g}^{i}}\left( \left. Z_{{{m}^{i}}}^{i} \right|{{{\pmb X}}_{n}}, {{u}^{i}} \right)}{{\pi }_+}\left( {{{\pmb X}}_{n}} \right)}{\rm d}{{x}_{1:n}}{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \end{align} $$ (21) $$ \begin{align} {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) = \, &\frac{1}{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}\times\\&\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{n}}}}}} {\prod\limits_{i = 1}^{s}{{{g}^{i}}\left( \left. Z_{{{m}^{i}}}^{i} \right|{{{\pmb X}}_{n}}, {{u}^{i}} \right)}{{\pi }_+}\left( {{{\pmb X}}_{n}} \right)}{\rm d}{{x}_{1:n}}{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \end{align} $$ (22) $$ \begin{align} &{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) = \\ &n!{{\rm e}^{-\sum\nolimits_{i = 1}^{s}{{{\lambda }^{i}}}}}\prod\limits_{i = 1}^{s}{{{\left[ {{\lambda }^{i}} \right]}^{{{m}^{i}}}}}\sum\limits_{(I, \xi )\in {{\mathcal{F}}_{n}}(\mathbb{L})\times \Xi }{\omega _+^{(I, \xi )}}\sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}} {{{\left\langle p_+^{(\xi )}\left( \cdot , \ell \right), \prod\limits_{i = 1}^{s}{{{\delta }_{{{({{\theta }^{i}})}^{-1}}(\{0:{{m}^{i}}\})}}(I)\varphi _{{{u}^{i}}}^{i}\left( \cdot , \ell ;{{\theta }^{i}} \right)} \right\rangle }^{I}}} \end{align} $$ (23) $$ \begin{align} &{{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) = \\ &\frac{n!{{\rm e}^{-\sum\nolimits_{i = 1}^{s}{{{\lambda }^{i}}}}}}{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}\prod\limits_{i = 1}^{s}{{{\left[ \alpha _{{\hat{n}}}^{i}{{\lambda }^{i}} \right]}^{{{m}^{i}}}}}\sum\limits_{(I, \xi )\in {{\mathcal{F}}_{n}}(\mathbb{L})\times \Xi }{\omega _+^{(I, \xi )}}\sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}} {{{\left\langle p_+^{(\xi )}\left( \cdot , \ell \right), \prod\limits_{i = 1}^{s}{{{\delta }_{{{({{\theta }^{i}})}^{-1}}(\{0:{{m}^{i}}\})}}(I)\phi _{\hat{n}, {{u}^{i}}}^{i}\left( \cdot , \ell ;{{\theta }^{i}} \right)} \right\rangle }^{I}}} \end{align} $$ (24) 其中
$$ \begin{align} \begin{split} \varphi _{{{u}^{i}}}^{i}\left( x, \ell ;{{\theta }^{i}} \right)& = {{\delta }_{0}}\left( {{\theta }^{i}}(\ell ) \right)\left( 1-p_{D}^{i}\left( x, \ell , {{u}^{i}} \right) \right)+\\&\left( 1-{{\delta }_{0}}\left( {{\theta }^{i}}(\ell ) \right) \right)\frac{p_{D}^{i}\left( x, \ell , {{u}^{i}} \right)}{{{\lambda }^{i}}} \end{split} \end{align} $$ (25) $$ \begin{align} \begin{split} \phi _{\hat{n}, {{u}^{i}}}^{i}\left( x, \ell ;{{\theta }^{i}} \right)& = {{\delta }_{0}}\left( {{\theta }^{i}}(\ell ) \right)\left( 1-p_{D}^{i}\left( x, \ell , {{u}^{i}} \right) \right)+\\&\left( 1-{{\delta }_{0}}\left( {{\theta }^{i}}(\ell ) \right) \right)\frac{\beta _{{\hat{n}}}^{i}p_{D}^{i}\left( x, \ell , {{u}^{i}} \right)}{\alpha _{{\hat{n}}}^{i}{{\lambda }^{i}}} \end{split} \end{align} $$ (26) $$ \begin{align} \alpha _{{\hat{n}}}^{i} = \int_{\mathbb{Z}_{\hat{n}1}^{i}}{f_{\kappa }^{i}\left( {{z}^{i}} \right){\rm d}{{z}^{i}}} \end{align} $$ (27) $$ \begin{align} \beta _{{\hat{n}}}^{i} = \int_{\mathbb{Z}_{\hat{n}1}^{i}}{{{g}^{i}}\left( \left. {{z}^{i}} \right|x, \ell , {{u}^{i}} \right){\rm d}{{z}^{i}}} \end{align} $$ (28) 既然$ q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $关于$ {{{\pmb x}}_{1:n}} $排序不变, 那么它对于$ {{{\pmb x}}_{1:n}} $中任意变量的边缘密度均相同, 记为
$$ \begin{align} &{{q}_{n}}\left( \left. {\pmb x}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) = \\& \qquad\int_{{{\mathbb{X}}_{n-1}}}{q\left( \left. \left\{ {\pmb x}, {{{\pmb x}}_{2:n}} \right\}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right){\rm d}{{{\pmb x}}_{2:n}}} \end{align} $$ (29) 将式(20)代入式(29), 再根据假设2和恒等式$ {{\delta }_{n}}\left( \left| \left\{ \ell , {{\ell }_{2:n}} \right\} \right| \right) = {{\delta }_{n-1}}\left( \left| \left\{ {{\ell }_{2:n}} \right\} \right| \right)\left( 1-{{1}_{\left\{ {{\ell }_{2:n}} \right\}}}(\ell ) \right) $, $ {{q}_{n}}\left( \left. {\pmb x}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $可写为
$$ \begin{align} &{{q}_{n}}\left( \left. {\pmb x}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) = \frac{1}{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}\\&\qquad\sum\limits_{{{\ell }_{2:n}}\in {{\mathbb{L}}_{n-1}}}{{{\delta }_{n-1}}\left( \left| \left\{ {{\ell }_{2:n}} \right\} \right| \right)\left( 1-{{1}_{\left\{ {{\ell }_{2:n}} \right\}}}(\ell ) \right)}\\&\qquad\sum\limits_{(I, \xi )\in {{\mathcal{F}}_{n}}(\mathbb{L})\times \Xi }{\omega _+^{(I, \xi )}{{\delta }_{I}}\left( \left\{ \ell , {{\ell }_{2:n}} \right\} \right)} \\&\qquad \int_{{{\mathbb{X}}_{n-1}}}{\prod\limits_{i = 1}^{s}{{{g}^{i}}\left( \left. Z_{{{m}^{i}}}^{i} \right|\left\{ {\pmb x}, {{{\pmb x}}_{2:n}} \right\}, {{u}^{i}} \right)}p_+^{(\xi )}({\pmb x})}\\&\qquad\prod\limits_{t = 2}^{n}{p_+^{(\xi )}\left( {{{\pmb x}}_{t}} \right)}{\rm d}{{{\pmb x}}_{2:n}} \end{align} $$ (30) 在$ s $个传感器中, 假设第$ {{i}_{1}}\ne , \cdots, \ne {{i}_{r}} $个传感器收到了目标$ {\pmb x} $的观测, 第$ {{i}_{1+r}}\ne , \cdots, \ne {{i}_{s}} $个传感器未收到目标$ {\pmb x} $的观测, $ r = 0, 1, \cdots, s $, 其中$ r = 0 $/$ r = s $表示所有传感器均未/已收到目标$ {\pmb x} $的观测.将式(7)代入式(30)并化简后可得式(31).
定理1. 给定假设1~3, 式(14)所示的多传感器多目标MSE的下界为式(32).
$$ \begin{align} & {{q}_{n}}\left( \left. x, \ell , Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) = \frac{1}{{{\Omega }_{n, {{m}^{1:s}}}} \left( {{u}^{1:s}} \right)}{{\rm e}^{-\sum\nolimits_{i = 1}^{s} {{{\lambda }^{i}}}}}\prod\limits_{i = 1}^{s}{{{\left[ {{\kappa }^{i}} \right]}^{{{Z}^{i}}}}}\sum\limits_{(I, \xi )\in {{\mathcal{F}}_{n}}(\mathbb{L})\times \Xi }{{1}_{I}}(\ell ) \omega _+^{(I, \xi )}p_+^{(\xi )}(x, \ell )\times \\ &\qquad \sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}} {\sum\limits_{0\le {{i}_{1}}\ne , \cdots, \ne {{i}_{r}} \le s}{\sum\limits_{{{z}^{{{i}_{1}}}}\in Z_{{{m}^{{{i}_{1}}}}} ^{{{i}_{1}}}}{\cdots\sum\limits_{{{z}^{{{i}_{r}}}}\in Z_{{{m}^ {{{i}_{r}}}}}^{{{i}_{r}}}}{{}}}}}\prod\limits_{j = 1}^{r}{{{\delta }_{{{({{\theta }^ {{{i}_{j}}}})}^{-1}}(\{0:{{m}^{{{i}_{j}}}}-1\})}}\left( I-\{\ell \} \right) p_{D}^{{{i}_{j}}}\left( x, \ell , {{u}^{{{i}_{j}}}} \right)}\times \\ &\qquad{{g}^{{{i}_{j}}}}\left( \left. {{z}^{{{i}_{j}}}} \right|x, \ell , {{u}^{{{i}_{j}}}} \right)\prod \limits_{j = 1+r}^{s}{{{\delta }_{{{({{\theta }^{{{i}_{j}}}})}^{-1}} (\{0:{{m}^{{{i}_{j}}}}\})}}\left( I-\{\ell \} \right) \left( 1-p_{D}^{{{i}_{j}}}\left( x, \ell , {{u}^{{{i}_{j}}}} \right) \right)} \times\\ &\qquad \begin{matrix} {} \\ \end{matrix}{{\left\langle p_+^{(\xi )}\left( \cdot , \ell \right), \prod\limits_{j = 1}^{r}{\psi _{Z_{{{m}^{{{i}_{j}}}}}^ {{{i}_{j}}}-\{{{z}^{{{i}_{j}}}}\}, {{u}^{{{i}_{j}}}}}^{{{i}_{j}}} \left( \cdot , \ell ;{{\theta }^{{{i}_{j}}}} \right)}\cdot \prod \limits_{j = 1+r}^{s}{\psi _{Z_{{{m}^{{{i}_{j}}}}}^{{{i}_{j}}}, {{u}^{{{i}_{j}}}}}^{{{i}_{j}}}\left( \cdot , \ell ; {{\theta }^{{{i}_{j}}}} \right)} \right\rangle }^{I-\{\ell \}}} \end{align} $$ (31) $$ \begin{align} \begin{split} {{\underline{\sigma }}^{2}}\left( {{u}^{1:s}} \right)& = \sum\limits_{{{m}^{s}} = 0}^{\infty }{\cdots }\sum\limits_{{{m}^{1}} = 0}^{\infty }{\sum\limits_{n = 0}^{\infty }{\sum\limits_{\hat{n} = 0, n+\hat{n}>0}^{\infty }}}\\&{\frac{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right){{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}{{{m}^{1}}!\cdots {{m}^{s}}!n!}}\cdot\\&\left( {{\varepsilon }_{\hat{n}, n}}{{\Phi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)+(1-{{\varepsilon }_{\hat{n}, n}}){{c}^{2}} \right) \end{split} \end{align} $$ (32) 式中, $ c $表示OSPA的截断误差, $ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $分别如式(23)和(24)所示,
$$ \begin{align} \begin{split} &{{\Phi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) = \\&\min \left( {{c}^{2}}, \frac{1}{{{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}\sum\limits_{l = 1}^{L}{{{\left[ J_{\hat{n}, n, {{m}^{1:s}}}^{-1}\left( {{u}^{1:s}} \right) \right]}^{l, l}}} \right) \end{split} \end{align} $$ (33) $$ \begin{align} {{\varepsilon }_{\hat{n}, n}} = \frac{\min \left( \hat{n}, n \right)}{\max \left( \hat{n}, n \right)} \end{align} $$ (34) 其中, $ L $表示无标签状态$ x $的维数, $ {{J}_{\hat{n}, n, {{m}^{1:s}}}} \left( {{u}^{1:s}} \right) $表示给定$ \Big( |\hat{{\pmb X}}| = \hat{n}, |{\pmb X}| = n, |{{Z}^{1}}| = {{m}^{1}}, \cdots, $ $ |{{Z}^{s}}| = {{m}^{s}} \Big) $时关于$ {{u}^{1:s}} $的$ L\times L $维FIM,
$$ \begin{align} &{{\left[ {{J}_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) \right]}^{i, j}} = \frac{-1}{\varpi _{\hat{n}, n, {{m}^{1:s}}}^{2}\left( {{u}^{1:s}} \right)}\times\\&\qquad\int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{1}}}{{{q}_{n}}\left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)}}}\\&\qquad\frac{{{\partial }^{2}}\log {{q}_{n}}\left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)}{\partial {{x}^{i}}\partial {{x}^{j}}}{\rm d}x{\rm d}z_{1:{{m}^{1}}}^{1}\cdots dz_{1:{{m}^{s}}}^{s}, \\&\qquad\qquad\qquad i, j = 1, \cdots, L \end{align} $$ (35) 式(35)中, 令$ {{J}_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) = \infty $.若$ \mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}\bigcup $ $ \cdots \bigcup \mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s} = \emptyset $, $ \hat{n} = 0, 1, \cdots, $ $ \infty $; $ {{q}_{n}}\left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $为式(31)的无标签版, 即
$$ \begin{align} \begin{split} {{q}_{n}}\left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) = \sum\limits_{\ell \in {{\mathbb{L}}_{1}}}{{{q}_{n}}\left( \left. x, \ell , Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)} \end{split} \end{align} $$ (36) 定理1的证明见附录B.
说明1. 若正则条件成立且$ {{\partial }^{2}}\log {{q}_{n}} $ $ \left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)/{\partial {{x}^{i}}\partial {{x}^{j}}}\; $存在, 那么式(35)所示的FIM必存在.
说明2. $ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{J}_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $的计算公式中包含$ \mathbb{Z}_{\hat{n}, {{m}^{1:s}}}^{1:s} $上的积分且没有解析表达式, 故需要利用数值积分法[30]先对$ \mathbb{Z}_{\hat{n}, {{m}^{1:s}}}^{1:s} $内的点进行采样再计算获得.为了降低计算代价, 采样点可选为预测理想测量集(Predicted ideal measurement set, PIMS)[6, 17], 数值积分可选择效率较高的伪Monte Carlo (MC)积分法[31].
最终, 对于每一个$ {{u}^{1:s}}\in \mathbb{U}_{D}^{1:s} $, 表 1给出了定理1所示的多目标MSE界$ {{\underline{\sigma }}^{2}}\left( {{u}^{1:s}} \right) $的计算步骤.
表 1 本文多目标MSE界的计算步骤Table 1 The calculation steps of the multi-target MSE bound in this paper步骤 计算多目标MSE界 第1步: 根据式(17)所示的多目标$ \delta $-GLMB预测密度$ {{\pi }_+}({\pmb X}) $产生$ M $个采样集$ {{\Psi}_{S}} = \left\{\tilde{{\pmb X}}_+^{(1)}, \cdots, \tilde{{\pmb X}}_+^{(M)} \right\} $; 第2步: 对于每一个$ \tilde{{\pmb X}}_+^{(j)}\in {{\Psi }_{S}} $, 根据式(7)所示的似然函数$ {{g}^{i}}\left( \cdot \left| \tilde{{\pmb X}}_+^{(j)}, {{u}^{i}} \right. \right) $, $ i = 1, \cdots, s $, 产生s个传感器的PIMS$ \tilde{Z}_{{{m}^{1:s}}}^{1:s, (j)} $; 第3步: 根据式(18)和(19)将所有的PIMS$ \left\{ \tilde{Z}_{{{m}^{1:s}}}^{1:s, (j)} \right\}_{j = 1}^{M} $分别划分到各个测量子空间$ \mathbb{Z}_{0, {{m}^{1:s}}}^{1:s}, \mathbb{Z}_{1, {{m}^{1:s}}}^{1:s}, \cdots, \mathbb{Z}_{\infty , {{m}^{1:s}}}^{1:s} $; 第4步: 给定$ \mathbb{Z}_{\hat{n}, {{m}^{1:s}}}^{1:s} $所包含的PIMS, 对式(24)和(35)应用伪MC积分法可得$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{J}_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $; 第5步: 根据式(23)、(33)和(34)计算得到$ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $, $ {{\Phi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{\varepsilon }_{\hat{n}, n}} $; 第6步: 将所求得的$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $, $ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $, $ {{\varepsilon }_{\hat{n}, n}} $和$ {{\Phi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $代入式(32)可得$ {{\underline{\sigma }}^{2}}\left( {{u}^{1:s}} \right) $. 3. 多传感器控制的子优化算法
本节建议采用混合罚函数法[27]或复合形法[28]来寻找式(15)的子最优解.两者的计算量随传感器个数$ s $的增加近似呈多项式增长.
混合罚函数法. 对于式(15)所示的有约束优化问题, 通过构造如下的增广目标函数将其松弛为无约束优化问题, 其中$ r>0 $称为障碍因子, 惩罚项$ \sum\nolimits_{i = 1}^{l}{\gamma _{i}^{-1}\left( {{u}^{1:s}} \right)} $限制搜索跑出不等式约束确定的区域, 惩罚项$ \sum\nolimits_{j = 1}^{m}{\nu _{j}^{2}\left( {{u}^{1:s}} \right)} $迫使搜索点向等式约束面靠近.
$$ \begin{align} F\left( {{u}^{1:s}}, r \right) = \, &{{\underline{\sigma }}^{2}} \left( {{u}^{1:s}} \right)+r\sum\limits_{i = 1}^{l} {\gamma _{i}^{-1}\left( {{u}^{1:s}} \right)}+\\& \frac{1}{\sqrt{r}}\sum\limits_{j = 1}^{m}{\nu _{j}^{2} \left( {{u}^{1:s}} \right)} \end{align} $$ (37) 对于松弛后的问题, 可利用无约束优化方法进行求解, 此处建议选择坐标下降法[23]. 表 2给出了本文混合罚函数法的主要执行步骤.
表 2 混合罚函数法Table 2 Mixed penalty function method步骤 混合罚函数求解过程 第1步: 选定满足不等式约束的初始点$ u_{(0)}^{1:s} $, 初始障碍因子$ r $及其缩小系数$ 0<C<1 $, 令$ i = 0 $; 第2步: 从$ j = 1 $到$ j = s $依次求$ u_{(i + 1)}^j = \arg {\mkern 1mu} \;\mathop {\min }\limits_{{u^j}} F\left( {u_{(i + 1)}^1, \cdots ,u_{(i + 1)}^{j - 1},{u^j},u_{(i)}^{j + 1}, \cdots ,u_{(i)}^s,r} \right) $, 此处只有$ {{u}^{j}} $为变量其余均为常量; 第3步: 若$ \sum\nolimits_{j = 1}^{s}{\left\| u_{(i+1)}^{j}-u_{(i)}^{j} \right\|}\le {{\varepsilon }_{2}} $, 转第4步; 否则令$ i = i+1 $, 转第2步; 第4步: 若$ \sum\nolimits_{j=1}^{s}{\left\| u_{(i+1)}^{j}-u_{(0)}^{j} \right\|}\le {{\varepsilon }_{1}} $, 输出$ u_{(i+1)}^{1:s} $为式(15)的解; 否则令$ r=Cr $, $ u_{(0)}^{1:s}=u_{(i+1)}^{1:s} $, $ i=0 $, 转第2步. 为了尽可能地使混合罚函数法收敛到全局最优点并加快收敛速度, 可采用文献[27]建议的方法选取表 2中的初始点$ u_{(0)}^{1:s} $, 初始障碍因子$ r $及其缩小系数$ C $.
复合形法. 该方法要求仅有不等式约束.
记$ \vec{u} = {{\left[ {{\left[ {{u}^{1}} \right]}^{\text{T}}}, \cdots, {{\left[ {{u}^{s}} \right]}^{\text{T}}} \right]}^{\text{T}}} $为$ s $个传感器的全体控制矢量, $ {{\vec{\mathbb{U}}}_{D}} $为$ \vec{u} $的可行域, $ N $为$ \vec{u} $的维数. 表 3给出了复合形法的主要执行步骤.
表 3 复合形法Table 3 Complex method步骤 复合形法求解过程 第1步: 选定$ {{\vec{u}}_{1}}, \cdots, {{\vec{u}}_{k}}\in{{\vec{\mathbb{U}}}_{D}} $ $ (N+2\le k\le2N) $构成初始复合形的$ k $个顶点, 为了避免降维, 要求至少有$ N+1 $个点的连线所形成的矢量有$ N $个是线性独立的; 第2步: 找出$ k $个顶点中目标函数值最大的点(最坏点)$ {{\vec{u}}_{H}} $和最小的点(最好点) $ {{\vec{u}}_{L}} $, $ {{\underline{\sigma }}^{2}} \left( {{{\vec{u}}}_{H}} \right) = \max\left\{ {{{\underline{\sigma }}}^{2}} \left( {{{\vec{u}}}_{1}}\right), \cdots, {{{\underline{\sigma }}}^{2}} \left({{{\vec{u}}}_{k}} \right) \right\} $, $ {{\underline{\sigma}}^{2}}\left( {{{\vec{u}}}_{L}} \right) = \min \left\{{{{\underline{\sigma }}}^{2}}\left( {{{\vec{u}}}_{1}}\right), \cdots, {{{\underline{\sigma }}}^{2}}\left({{{\vec{u}}}_{k}}\right) \right\} $; 第3步: 求除去$ {{\vec{u}}_{H}} $外其余$ k-1 $个顶点的中心点$ {{\vec{u}}_{0}} = \frac{1}{k-1}\left( \sum\limits_{i = 1}^{k}{{{{\vec{u}}}_{i}}}-{{{\vec{u}}}_{H}} \right) $.若$ {{\vec{u}}_{0}}\in{{\vec{\mathbb{U}}}_{D}} $, 转第4步; 否则, 转第1步, 重新选择初始顶点; 第4步: 以$ {{\vec{u}}_{0}} $为轴心求$ {{\vec{u}}_{H}} $的映射点$ {{\vec{u}}_{R}} = {{\vec{u}}_{0}}+t\left({{{\vec{u}}}_{0}}- {{{\vec{u}}}_{H}} \right) $, 其中$ t>0 $称为映射系数, 一般先取$ t = 1.3 $.若$ {{\vec{u}}_{R}}\in{{\vec{\mathbb{U}}}_{D}} $, 转第5步; 否则, 令$ t = 0.5t $, 转第4步; 第5步: 若$ {{\underline{\sigma }}^{2}}\left({{{\vec{u}}}_{R}}\right)<{{\underline{\sigma }}^{2}}\left( {{{\vec{u}}}_{H}} \right) $, 令$ {{\vec{u}}_{H}} = {{\vec{u}}_{R}} $, 转第6步; 否则, 令$ t = 0.5t $,转第4步.若$ t $已很小(如小于$ {{10}^{-5}} $)还不能满足上式, 说明由$ {{\vec{u}}_{H}} $和$ {{\vec{u}}_{0}} $形成的映射方向不利.为改变映射方向, 令$ {{\vec{u}}_{H}} = {{\vec{u}}_{SH}} $, 其中$ {{\vec{u}}_{SH}} $为次坏点, 转第3步; 第6步: 若$ {{\left( \frac{1}{k}\sum\limits_{i=1}^{k}{{{\left[ {{{\underline{\sigma }}}^{2}}\left( {{{\vec{u}}}_{0}}\right)-{{{\underline{\sigma }}}^{2}}\left( {{{\vec{u}}}_{i}}\right)\right]}^{2}}} \right)}^{\frac{1}{2}}}\le {{\varepsilon }_{1}} $或$ \mathop {\max }\limits_{1 \le i \le k} {\mkern 1mu} \left\| {{{\vec u}_i} - {{\vec u}_0}} \right\| \le {\varepsilon _2} $, 输出$ {{\vec{u}}_{L}} $为式(15)的解; 否则, 转第2步. 表 3中复合形的初始顶点$ {{\vec{u}}_{1}}, \cdots, {{\vec{u}}_{k}} $可参照文献[28]的方法选定.为了尽可能地使复合形法收敛到全局最优点, Krus等[32]对映射点$ {{\vec{u}}_{R}} $的计算公式进行了改进, 但会增加一定的收敛时间.
综上, 本文多传感器控制及多目标跟踪算法的执行流程如表 4所示.
表 4 本文多传感器控制及多目标跟踪算法的执行流程Table 4 The implementation process of multi-sensor control and multi-target tracking algorithm proposed in this paper第1步: 对于式(15)所示的优化问题, 采用穷尽搜索、混合罚函数或复合形法三者之一计算多传感器最优控制量$ {{\left[ {{u}^{1:s}} \right]}^{*}} $; 第2步: 执行控制命令$ {{\left[ {{u}^{1:s}} \right]}^{*}} $后, 可获得真实测量集$ {{Z}^{1:s}} $; 第3步: 将$ {{Z}^{1:s}} $代入$ \delta $-GLMB滤波器的更新步, 可得多目标$ \delta $-GLMB后验密度$ \pi \left( \left.{\pmb X} \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) $; 第4步: 根据假设3可得多目标状态估计集$ {{\hat{{\pmb X}}}_{{\hat{n}}}} $; 第5步: 将$ \pi \left( \left. {\pmb X} \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) $代入$ \delta $-GLMB滤波器的预测步, 可得下一时刻的多目标$ \delta $-GLMB预测密度$ {{\pi }_+}({\pmb X}) $.转第1步. 最后, 为了详细说明建议的多传感器控制算法和已有的基于GCI融合准则的多传感器控制算法[19-20]的区别, 我们在表 5中对两者进行了对比.
表 5 本文多传感器控制算法与基于GCI的多传感器控制算法的比较Table 5 A comparison between multi-sensor control algorithm proposed in this paper and multi-sensor control algorithm based on GCI本文多传感器控制算法 基于GCI的多传感器控制算法 目标函数 多目标MSE界 CS散度或PEECS 滤波器 $ \delta $-GLMB 边缘$ \delta $-GLMB或LMB (两者均为$ \delta $-GLMB的近似) 融合结构 集中式 分布式 通讯量 需传递各个传感器的测量集 需传递各个传感器的局部后验密度 计算方式 并行处理 并行处理 表 5中GCI准则为:给定局部多目标密度$ {{\pi }^{i}}({\pmb X}) $和归一化权重$ {{\upsilon }^{i}} $, $ i\in \mathcal{I} $, 使得Kullback-Leiblerb (KL)散度最小的融合密度如式(38)所示.
在利用GCI准则前, 其归一化融合权重$ {{\upsilon }^{i}} $, $ i\in \mathcal{I} $必须提前设定好.
$$ \begin{align} \bar{\pi }({\pmb X}) = \frac{\prod\nolimits_{i\in \mathcal{I}}{{{\left[ {{\pi }^{i}}({\pmb X}) \right]}^{{{\upsilon }^{i}}}}}}{\int{\prod\nolimits_{i\in \mathcal{I}}{{{\left[ {{\pi }^{i}}({\pmb X}) \right]}^{{{\upsilon }^{i}}}}}\delta {\pmb X}}} \end{align} $$ (38) 本文方法的主要贡献是可以避免由于GCI融合权重设定不当而引起的多传感器控制效率低下以及最终的多目标跟踪精度达不到预期效果.尽管如此,基于集中式结构的多传感器控制方法在并行性、灵活性以及抗打击能力等方面不如基于分布式结构的多传感器控制方法.这是两种处理结构的固有特性所决定的.
4. 仿真分析
4.1 例1.传感器个数较少的场景
在二维区域$ \mathcal{S} = [-50\, \text{m}, 50\, \text{m}]\times [-50\, \text{m}, 50\, \text{m}] $中共有$ s = 4 $个运动可控的传感器对多个目标进行观测, 观测时长$ T = 30 $步.单目标状态记为$ {\pmb x} = \left( x, \ell \right) $, 标签$ \ell = \left( {{k}_{B}}, {{i}_{B}} \right) $, 其中$ {{k}_{B}} $为该目标出生时刻, $ {{i}_{B}} $是同一时刻区分出生目标的索引; 运动状态$ x = {{\left[ {{p}_{x}}, {{{\dot{p}}}_{x}}, {{p}_{y}}, {{{\dot{p}}}_{y}}, w \right]}^{\text{T}}} $, 其中$ \left( {{p}_{x}}, {{p}_{y}} \right) $和$ \left( {{{\dot{p}}}_{x}}, {{{\dot{p}}}_{y}} \right) $分别为$ X $和$ Y $方向的位置和速度, $ w $为转弯率.目标运动满足转弯模型[33], 其转移密度为如下高斯形式:
$$ \begin{align} f\left( \left. {{x}_+}, {{\ell }_+} \right|x, \ell \right) = \mathcal{N}\left( {{x}_+};u(x), Q \right){{\delta }_{\ell }}({{\ell }_+}) \end{align} $$ (39) 式中, $ u(\cdot ) $和$ Q $分别为状态转移函数和过程噪声协方差矩阵,
$$ \begin{align} u(x) = \left[ \begin{matrix} {{p}_{x}}+{{{\dot{p}}}_{x}}\frac{\sin (w\Delta )}{w}-{{{\dot{p}}}_{y}}\frac{1-\cos (w\Delta )}{w} \\ {{{\dot{p}}}_{x}}\cos (w\Delta )-{{{\dot{p}}}_{y}}\sin (w\Delta ) \\ {{p}_{y}}+{{{\dot{p}}}_{x}}\frac{1-\cos (w\Delta )}{w}+{{{\dot{p}}}_{y}}\frac{\sin (w\Delta )}{w} \\ {{{\dot{p}}}_{x}}\sin (w\Delta )+{{{\dot{p}}}_{y}}\cos (w\Delta ) \\ w{{\rm e}^{-{\Delta }/{{{\tau }_{w}}}\;}} \\ \end{matrix} \right] \end{align} $$ (40) $$ \begin{align} Q = \left[ \begin{matrix} \frac{{{\Delta }^{4}}}{4}q_{x}^{2} & \frac{{{\Delta }^{3}}}{2}q_{x}^{2} & {0} & {0} & {0} \\ \frac{{{\Delta }^{3}}}{2}q_{x}^{2} & \frac{{{\Delta }^{2}}}{2}q_{x}^{2} & {0} & {0} & {0} \\ {0} & {0} & \frac{{{\Delta }^{4}}}{4}q_{y}^{2} & \frac{{{\Delta }^{3}}}{2}q_{y}^{2} & {0} \\ {0} & {0} & \frac{{{\Delta }^{3}}}{2}q_{y}^{2} & \frac{{{\Delta }^{2}}}{2}q_{y}^{2} & {0} \\ {0} & {0} & {0} & {0} & q_{w}^{2} \\ \end{matrix} \right] \end{align} $$ (41) 其中, $ \Delta $为采样间隔, $ {{\tau }_{w}} $为转弯率的时间相关常数, $ {{q}_{x}} $和$ {{q}_{y}} $分别为$ X $和$ Y $方向的加速度, $ {{q}_{w}} $为转弯率的噪声标准差.本例中令$ \Delta = 1\, \text{s} $, $ {{\tau }_{w}} = 20\, \text{s} $, $ {{q}_{x}} = 0.1\, \text{m/}{{\text{s}}^{\text{2}}} $, $ {{q}_{y}} = 0.05\, \text{m/}{{\text{s}}^{\text{2}}} $, $ {{q}_{w}} = 0.01\, \text{rad/s} $, 目标存活概率为$ {{p}_{S}}(x, \ell ) = 0.95 $.
新生目标建模为一个标签泊松RFS, 强度为
$$ \begin{align} {{D}_{B}}(x) = \sum\limits_{i = 1}^{5}{0.04\mathcal{N}\left( x;{{x}_{B, i}}, {{Q}_{B}} \right)} \end{align} $$ (42) 其中, $ {{x}_{B, i}} $和$ {{Q}_{B}} $分别为新生目标状态均值和协方差阵.本例中令
$$ \begin{align*} {{x}_{B, 1}} = \, &{{[30\, \text{m}, 1\, \text{m/s}, -30\, \text{m}, 2\, \text{m/s}, 0.2\, \text{rad/s}]}^{\text{T}}} \\ {{x}_{B, 2}} = \, &{{[30\, \text{m}, -5\, \text{m/s}, 30\, \text{m}, -3\, \text{m/s}, 0.2\, \text{rad/s}]}^{\text{T}}} \\{{x}_{B, 3}} = \, &{{[-30\, \text{m}, -3\, \text{m/s}, -30\, \text{m}, 4\text{m/s}, -0.3\, \text{rad/s}]}^{\text{T}}} \\{{x}_{B, 4}} = \, &{{[-30\, \text{m}, 1\, \text{m/s}, 30\, \text{m}, -3\, \text{m/s}, 0.1\, \text{rad/s}]}^{\text{T}}} \\{{x}_{B, 5}} = \, &{{[0\, \text{m}, -4\, \text{m/s}, 0\, \text{m}, 4\text{m/s}, -0.3\, \text{rad/s}]}^{\text{T}}} \\{{Q}_{B}} = \, &{\rm diag}\{25\, {{\text{m}}^{\text{2}}}, 0.1\, {{\text{m}}^{2}}\text{/}{{\text{s}}^{\text{2}}}, 25\, {{\text{m}}^{\text{2}}}, 0.1\, {{\text{m}}^{2}}\text{/}{{\text{s}}^{\text{2}}}, \\& 0.01\, \text{ra} {{\text{d}}^{2}}\text{/}{{\text{s}}^{2}}\} \end{align*} $$ 其中, diag$ \{\cdot\} $表示对角矩阵.
传感器初始位置分别设为$ u_{0}^{1} = {{[45\, \text{m}, 45\, \text{m}]}^{\text{T}}} $, $ u_{0}^{2} = {{[-45\, \text{m}, -45\, \text{m}]}^{\text{T}}} $, $ u_{0}^{3} = {{[45\, \text{m}, -45\, \text{m}]}^{\text{T}}} $, $ u_{0}^{4} = {{[-45\, \text{m}, 45\, \text{m}]}^{\text{T}}} $.若传感器$ i $当前位置为$ {{u}^{i}} = {{\left[ p_{x, u}^{i}, p_{y, u}^{i} \right]}^{\text{T}}} $, 在控制命令作用下它下一时刻的可能位置集$ \mathbb{U}_+^{i} $为
$$ \begin{align} \mathbb{U}_+^{i} = \, &\Big[ p_{x, u}^{i}+\!j{{\rho }_{0}}\cos \left( k\frac{2\pi }{{{N}_{\theta }}} \right), p_{y, u}^{i}+\\&j{{\rho }_{0}}\sin\left( k\frac{2\pi }{{{N}_{\theta }}} \right) \Big]^{\text{T}}; \\&j = 0, \cdots, {{N}_{\rho }};k = 1, \cdots, {{N}_{\theta }} \end{align} $$ (43) 本例中令$ {{N}_{\rho }} = 2 $, $ {{N}_{\theta }} = 8 $和$ {{\rho }_{0}} = 5\, \text{m} $.若传感器运动出了区域$ \mathcal{S} $, 将其误差界设为$ \infty $.
4个传感器测量同步且传感器$ i $的单目标似然函数为如下高斯形式
$$ \begin{align} & {{g}^{i}}\left( \left. {{z}^{i}} \right|x, \ell , {{u}^{i}} \right) = \mathcal{N}\left( {{z}^{i}};{{h}^{i}}\left( x, {{u}^{i}} \right), {{R}^{i}}\left( x, {{u}^{i}} \right) \right), \\& \hskip40mm i = 1, \cdots, 4 \end{align} $$ (44) 式中, $ {{h}^{i}}\left( x, {{u}^{i}} \right) $和$ {{R}^{i}}\left( x, {{u}^{i}} \right) $分别为观测函数和观测噪声协方差阵, 两者均为目标运动状态$ x $和传感器位置$ {{u}^{i}} $的非线性函数.
4个传感器的观测性能显著不同, 具体体现在:
1) 观测函数不同.传感器1和2可同时获得目标距离和角度测量, 其观测函数为
$$ {h^i}\left( {x,{u^i}} \right) = \left[ {\begin{array}{*{20}{c}} {\left\| {x - {u^i}} \right\|}\\ {\arctan \frac{{{p_y} - p_{y,u}^i}}{{{p_x} - p_{x,u}^i}}} \end{array}} \right],{\rm{ }}\;i = 1,2$$ (45) 其中, 令$ \left\| x-{{u}^{i}} \right\| = \sqrt{{{\left( {{p}_{x}}-p_{x, u}^{i} \right)}^{2}}+{{\left( {{p}_{y}}-p_{y, u}^{i} \right)}^{2}}} $表示传感器$ i $与目标间的距离.
传感器3只能获得目标距离测量, 其观测函数为
$$ \begin{align} {{h}^{3}}\left( x, {{u}^{3}} \right) = \left\| x-{{u}^{3}} \right\| \end{align} $$ (46) 传感器4只能获得目标角度测量, 其观测函数为
$$ \begin{align} {{h}^{4}}\left( x, {{u}^{4}} \right) = \arctan \frac{{{p}_{y}}-p_{y, u}^{4}}{{{p}_{x}}-p_{x, u}^{4}} \end{align} $$ (47) 2) 测量噪声不同. 4个传感器的测量噪声均随目标与传感器间距离的增大而线性增长, 测量噪声协方差可分别写为
$$ \begin{align} &{{R}^{1}}\left( x, {{u}^{1}} \right) = \\&\left[ \begin{matrix} {{\left( 1+0.1\left\| x-{{u}^{1}} \right\| \right)}^{2}} & {0} \\ {0} & {{\left( 0.02+0.002\left\| x-{{u}^{1}} \right\| \right)}^{2}}\\ \end{matrix} \right] \end{align} $$ (48) $$ \begin{align} &{{R}^{2}}\left( x, {{u}^{2}} \right) = \\&\left[ \begin{matrix} {{\left( 5+0.02\left\| x-{{u}^{2}} \right\| \right)}^{2}} & {0}\\ {0} & {{\left( 0.1+0.0004\left\| x-{{u}^{2}} \right\| \right)}^{2}}\\ \end{matrix} \right] \end{align} $$ (49) $$ \begin{align} {{R}^{3}}\left( x, {{u}^{3}} \right) = {{\left( 2.5+0.05\left\| x-{{u}^{3}} \right\| \right)}^{2}} \end{align} $$ (50) $$ \begin{align} {{R}^{4}}\left( x, {{u}^{4}} \right) = {{\left( 0.05+0.001\left\| x-{{u}^{4}} \right\| \right)}^{2}} \end{align} $$ (51) 由以上四式可见, 传感器1的近距离观测精度最高, 但其随传感器与目标间距离增大而衰退最快; 相反, 传感器2的近距离观测精度最低, 但其随传感器与目标间距离增大而衰退最慢.
3) 检测概率不同. 4个传感器的检测概率均随目标与传感器间距离的增大而线性下降.
$$ \begin{align} p_{D}^{1}\left( x, \ell , {{u}^{1}} \right) = 0.98\left( 1-\text{0.007}\left\| x-{{u}^{1}} \right\| \right) \end{align} $$ (52) $$ \begin{align} p_{D}^{2}\left( x, \ell , {{u}^{2}} \right) = 0.8\left( 1-\text{0.001}\left\| x-{{u}^{2}} \right\| \right) \end{align} $$ (53) $$ \begin{align} p_{D}^{3}\left( x, \ell , {{u}^{3}} \right) = 0.92\left( 1-\text{0.005}\left\| x-{{u}^{3}} \right\| \right) \end{align} $$ (54) $$ \begin{align} p_{D}^{4}\left( x, \ell , {{u}^{4}} \right) = 0.86\left( 1-\text{0}\text{.003}\left\| x-{{u}^{4}} \right\| \right) \end{align} $$ (55) 4) 杂波密度不同. 4个传感器的泊松杂波强度为
$$ \begin{align} \begin{matrix} {{\kappa }^{i}}\left( {{z}^{i}} \right) = {{\lambda }^{i}}\mathcal{U}\left( {{z}^{i}};\mathcal{S} \right), & i = 1, \cdots, 4 \\ \end{matrix} \end{align} $$ (56) 其中, $ \mathcal{U}\left( \cdot ;\mathcal{S} \right) $为区域$ \mathcal{S} $上的均匀密度.令$ {{\lambda }^{1}} = 50 $, $ {{\lambda }^{2}} = 40 $, $ {{\lambda }^{3}} = 30 $, $ {{\lambda }^{4}} = 20 $.
传感器移动需要耗费能量, 其与移动距离成正比, 且每次控制所耗费的总能量不能超过门限$ E $, 故该能耗不等式约束模型可记为
$$ \begin{align} {{\gamma }_{1}}\left( {{u}^{1:s}} \right) = E -\sum\limits_{i = 1}^{s}{{{\chi }^{i}}\left\| {{u}^{i}}-u_{-}^{i} \right\|}\ge 0 \end{align} $$ (57) 其中, $ {{\chi }^{i}} $为传感器$ i $移动单位距离所耗费的能量.本例中设$ E = 100\, \text{J} $, 令$ {{\chi }^{1}} = 10\, {\text{J}}/{\text{m}} $, $ {{\chi }^{2}} = 8\, {\text{J}}/{\text{m}} $, $ {{\chi }^{3}} = 6\, {\text{J}}/{\text{m}} $, $ {{\chi }^{4}} = 4\, {\text{J}}/{\text{m}} $.
为了避免传感器自身以及传感器与目标间的相互碰撞, 还增加了如下的避撞不等式约束模型:
$$ \begin{align} {{\gamma }_{2}}\left( {{u}^{1:s}} \right) = \underset{1\le i\ne j\le s}{\mathop{\min }}\, \left\| {{u}^{i}}-{{u}^{j}} \right\|-{{T}_{1}}\ge 0 \end{align} $$ (58) $$ \begin{align} {{\gamma }_{3}}\left( {{u}^{1:s}} \right) = \underset{1\le i\le s, 1\le j\le {{{\hat{n}}}_+}}{\mathop{\min }}\left\| {{u}^{i}}-{{{\hat{x}}}_{+, j}} \right\|-{{T}_{2}}\ge 0 \end{align} $$ (59) 上两式的含义是保证任意两个传感器间以及任意传感器与任意预测目标间的最小距离分别大于各自的避撞门限距离$ {{T}_{1}} $和$ {{T}_{2}} $, 其中预测目标的个数$ {{\hat{n}}_+} $和运动状态$ {{\hat{x}}_{+, j}} $ ($ j = 1, \cdots, {{\hat{n}}_+} $)可根据假设3由式(17)所示的多目标先验密度$ {{\pi }_+}({\pmb X}) $获得.本例令$ {{T}_{1}} = {{T}_{2}} = 5\, \text{m} $.
本例中设每帧最大目标数为10, 每个传感器每帧收到的最大测量数为200, 截断距离$ c = 100 $.采用粒子滤波器[34]执行本算法, 平均采用1 000个粒子近似每个$ \delta $-GLMB RFS的航迹密度$ {{p}^{(\xi )}}\left( \cdot , \ell \right) $.由于$ \delta $-GLMB滤波器的历史关联映射项$ \xi $的个数随时间呈指数增长, 为了控制计算量将权重$ {{\omega }^{(I, \xi )}} $小于$ {{10}^{-5}} $的项删除.
把将要比较的方法分别简记为: CS散度+穷尽法[19]、误差界+穷尽法、误差界+混合罚函数法和误差界+复合形法.依照惯例将CS散度法的GCI权重设为相同.用于测试算法的台式机CPU为i7-8700K, 内存为32 GB.
首先将CS散度+穷尽法和本文的误差界+穷尽法进行比较.
图 1展示了单次实验中分别采用两种方法的传感器运动轨迹.
图 1 单次实验中CS散度+穷尽法和误差界+穷尽法的传感器运动轨迹(黑实线表示目标运动轨迹, $ \bigcirc $和$ \Delta $分别表示目标的起点和终点; 灰色实线表示传感器运动轨迹, $ \square $和其上方数字分别表示传感器位置和处于该位置的时刻, $ \blacksquare $和$ \boxplus $分别表示传感器的起点和终点)Fig. 1 Sensor trajectories in a simulation by using CS divergence with exhaustive search method and error bound with exhaustive search method(The black line is the target trajectory, $ \bigcirc $ and $ \Delta $ are the target starting point and ending point; the gray line is the sensor trajectory, $ \square $ and the number above it are the sensor position and the time when the sensor is located at the position, $ \blacksquare $ and $ \boxplus $ are the sensor starting point and ending point)图 2展示了分别采用随机法、CS散度+穷尽法和本文误差界+穷尽法经500次MC仿真平均后得到的多目标位置估计的OSPA误差随时间变化曲线.
首先, 图 1表明CS散度+穷尽法和本文的误差界+穷尽法均使各个传感器接近相应的存活目标, 但前者传感器4运动的距离最长, 传感器1运动的距离最短; 相反, 后者传感器1运动的距离最长, 传感器4运动的距离最短.
从图 2又可以看出, 采用误差界+穷尽法得到的OSPA误差最小且最接近建议的多目标MSE界, 下降速度也最快; CS散度+穷尽法次之; 随机法的OSPA误差最大, 下降速度最慢.这说明本文误差界+穷尽法的多传感器控制效果要优于CS散度+穷尽法, 导致其多目标跟踪性能相比后者有明显提高, 后者又优于随机法.造成上述结果的主要原因是:
1) 在CS散度+穷尽法中, GCI权重设为相同表示其认为各个传感器对于目标函数(多传感器多目标CS散度)具有相同影响.故在给定能耗约束下传感器4由于移动每米能耗最小而具有最高控制级, 最终移动距离最长; 相反传感器1由于移动每米能耗最大而具有最低控制级, 最终移动距离最短.
2) 误差界+穷尽法避免了GCI准则, 由定理1可知各个传感器对目标函数的影响与其自身似然函数的特性有关.这样即使传感器1移动每米能耗最大, 但由于其似然函数对该界的影响最大而具有最高控制级, 最终移动距离最长; 相反传感器4由于其似然函数对该界的影响最小而具有最低控制级, 最终移动距离最短.
3) 既然误差界+穷尽法相比CS散度+穷尽法对传感器1的控制力更强, 最终如图 2所示其具有最优的多目标跟踪精度.而随机法没有使传感器接近目标的能力, 故其跟踪精度最差.
4.2 例2:传感器个数较多的场景
将例1传感器个数$ s $和能耗约束门限$ E $分别变为($ s = 8, E = 200\, \text{J} $)、($ s = 12, E = 300\, \text{J} $)、($ s = 16, E = 400\, \text{J} $)和($ s = 20, E = 500\, \text{J} $).采用算法单次运行所耗费的CPU时间来衡量其计算量. 表 6展示了每种场景(包括例1)下4种方法经500次MC仿真平均后的OSPA误差终值和CPU耗时.
表 6 500次MC仿真平均后的OSPA误差终值(m)和CPU耗时(s)Table 6 500 MC run averages of final OSPA error distance (m) and CPU processing time (s)传感器控制算法 s = 4, E = 100J s = 4, E = 100J s = 4, E = 100J s = 4, E = 100J s = 4, E = 100J CS散度+穷尽法 19.5m/22s 13.9m/153s 9.2m/1062s 5.8m/7259s 4.6m/50973s 误差界+穷尽法 8.7m/23s 6.0m/162s 4.6m/1128s 3.9m/7698s 3.6m/53168s 误差界+混合罚函数法 8.9m/20s 6.3m/91s 5.1m/204s 4.5m/365s 4.1m/578s 误差界+复合形法 9.0m/16s 6.4m/81s 5.1m/206s 4.4m/573s 4.0m/1296s 建议的多目标MSE界 8.0m 5.4m 4.1m 3.5m 3.2m 从表 6可以看出, 随着传感器个数的增加,
1) 4种方法的OSPA误差以及建议的多目标MSE界均有所下降.误差界+穷尽法总保持最优且最接近建议的界, CS散度+穷尽法总是最差.尽管误差界+混合罚函数和误差界+复合形这两种子最优的方法当传感器数目较多时陷入局部极值点的可能性增大, 导致它俩相比误差界+穷尽法的OSPA误差增幅会稍有提高, 但仍与后者相接近.然而, 当传感器数目多到一定程度时, 4种方法跟踪精度的提高幅度变得相对较小, 并且它们的OSPA值也相差不大, 此时可理解为多目标跟踪性能的提高达到“饱和”状态.
2) 在计算时间方面, 误差界+混合罚函数法的增长速度最慢, 其次为误差界+复合形法, 这两种子最优的方法增长速度又明显慢于CS散度+穷尽法和误差界+穷尽法, 后两者的计算时间基本相差不大.当传感器个数较少时, 误差界+复合形法的计算时间最短, 误差界+混合罚函数法略次之, 但两种子最优的方法相对两种穷尽法计算速度提高并不明显; 当传感器个数较多时, 两种子最优方法的计算速度显著快于两种穷尽法, 且误差界+混合罚函数法超过误差界+复合形法成为计算时间最短的算法.这是由于传感器个数较多导致所优化的控制矢量维数也较高, 此时复合形法的收敛时间会明显增加.
5. 结论与展望
本文提出了一种带约束条件的针对多目标跟踪系统的多传感器控制算法.该算法以多目标MSE界作为传感器控制的代价函数.采用2阶OSPA测度定义多目标状态集和估计集的误差, 并将多目标Bayes递推建模为一个$ \delta $-GLMB RFS.为了降低计算代价, 建议采用混合罚函数法和复合形法这两种子最优的算法来替代穷尽搜索法求解该有约束的优化问题.仿真结果表明对于观测性能相差较大且有约束存在的多传感器控制系统, 本方法的多目标跟踪精度明显优于CS散度法等基于GCI准则的多传感器控制算法.并且当传感器个数较多时, 采用所建议的子最优算法相比采用穷尽搜索法, 其计算时间显著缩短而最终的多目标跟踪精度损失很小.
后续工作将聚焦于:
1) 当监控区域内存在大数量的传感器时, 需要将传感器选择和控制问题统一研究.
2) 将本文现有成果扩展到分布式处理结构.
附录A
应用条件概率公式, 后验概率$ {\rm P}\left( \left. |{\pmb X}| = n \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) $可写为
$$ \begin{align} {\rm P}\left( \left. |{\pmb X}| = n \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) = \frac{{\rm P}\left( \left. |{\pmb X}| = n, {{Z}^{1:s}} \right|{{u}^{1:s}} \right)}{{\rm P}\left( \left. {{Z}^{1:s}} \right|{{u}^{1:s}} \right)} \end{align} $$ (A1) 式中, 分子P$ \left( \left. |{\pmb X}| = n, {{Z}^{1:s}} \right|{{u}^{1:s}} \right) $表示以$ {{u}^{1:s}} $为条件的$ \left( |{\pmb X}| = n, {{Z}^{1:s}} \right) $的联合概率, 它可由下式获得:
$$ \begin{align} & {\rm P}\left( \left. |{\pmb X}| = n, {{Z}^{1:s}} \right|{{u}^{1:s}} \right) = \\&\qquad\frac{1}{n!}\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\int_{{{\mathbb{X}}_{n}}}{f\left( \left. {{{\pmb X}}_{n}}, {{Z}^{1:s}} \right|{{u}^{1:s}} \right){\rm d}{{x}_{1:n}}}} = \\ &\qquad \frac{1}{n!}\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}} {\int_{{{\mathbb{X}}_{n}}}{\prod\limits_{i = 1}^{s}{{{g}^{i}} \left( {{Z}^{i}}\left| {{{\pmb X}}_{n}}, {{u}^{i}} \right. \right)} {{\pi }_{+}}\left( {{{\pmb X}}_{n}} \right){\rm d}{{x}_{1:n}}}} \end{align} $$ (A2) 将假设2和式(7)代入式(61)并积分后可得:
$$ \begin{align} & P\left( \left. |{\pmb X}| = n, {{Z}^{1:s}} \right|{{u}^{1:s}} \right) = \\& \qquad\frac{1}{n!}\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\sum\limits_{\xi \in \Xi }{\Delta \left( {{{\pmb X}}_{n}} \right)\omega _+^{(I, \xi )}}} \sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}} \Bigg\langle p_+^{(\xi )}\left( \cdot , \ell \right), \\& \qquad \prod\limits_{i = 1}^{s}{{{\delta }_{{{({{\theta }^{i}})}^{-1}} (\{0:|{{Z}^{i}}|\})}}(I)\psi _{{{Z}^{i}}, {{u}^{i}}}^{i} \left( \cdot , \ell ;{{\theta }^{i}} \right)} \Bigg\rangle ^{I} = \\&\qquad \sum\limits_{(I, \xi )\in {{\mathcal{F}}_{n}}(\mathbb{L}) \times \Xi }{\omega _+^{(I, \xi )}} \sum\limits_{{{\theta }^{1:s}}\in {{\Theta }^{1:s}}} \Bigg\langle p_+^{(\xi )}\left( \cdot , \ell \right), \\& \qquad \prod\limits_{i = 1}^{s}{{{\delta }_{{{({{\theta }^{i}})}^{-1}} (\{0:|{{Z}^{i}}|\})}}(I)\psi _{{{Z}^{i}}, {{u}^{i}}}^{i} \left( \cdot , \ell ;{{\theta }^{i}} \right)} \Bigg\rangle ^{I} \end{align} $$ (A3) 上式中最后一行根据文献[7]的引理12得到.
根据全概率公式可将式(60)的分母$ P\left( \left. {{Z}^{1:s}} \right|{{u}^{1:s}} \right) $写为
$$ \begin{align} P\left( \left. {{Z}^{1:s}} \right|{{u}^{1:s}} \right) = \, & \sum\limits_{n = 0}^{\infty }{P\left( \left. |{\pmb X}| = n, {{Z}^{1:s}} \right|{{u}^{1:s}} \right)} = \\& \sum\limits_{(I, \xi )\in \mathcal{F} (\mathbb{L})\times \Xi }{\omega _+^{(I, \xi )}\sum\limits_{{{\theta }^{1:s}} \in {{\Theta }^{1:s}}}}\Bigg\langle p_+^{(\xi )} \left( \cdot , \ell \right), \\&\prod\limits_{i = 1}^{s} {{{\delta }_{{{({{\theta }^{i}})}^{-1}}(\{0:|{{Z}^{i}}|\})}}(I) \psi _{{{Z}^{i}}, {{u}^{i}}}^{i}\left( \cdot , \ell ;{{\theta }^{i}} \right)} \Bigg\rangle ^{I} \end{align} $$ (A4) MAP检测准则判定$ |\hat{{\pmb X}}| = \hat{n} $, 若
$$ \begin{align} \hat{n} = \arg \underset{n}{\mathop{\max }}\, \left\{ P\left( \left. |{\pmb X}| = n \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) \right\} \end{align} $$ (A5) 将式(62)和式(63)代入式(60)后再将所得结果代入式(64), 最终可得式(18)和式(19).
附录B
根据式(11)所示的标签RFS积分定义和式(20), 式(14)可重写为
$$ \begin{align} {{\sigma }^{2}}\left( {{u}^{1:s}} \right) = \, &\sum\limits_{{{m}^{s}} = 0}^{\infty }{\cdots \sum\limits_{{{m}^{1}} = 0}^{\infty }{\sum\limits_{n = 0}^{\infty }{\frac{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}{{{m}^{1}}!\cdots {{m}^{s}}!n!}}}}\times\\&\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\int_{\mathbb{Z}_{{{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{{{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{n}}}{q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)}}}}\times\\&{{e}^{2}}\left({{{\pmb X}}_{n}}, \hat{{\pmb X}}\right){\rm d}{{x}_{1:n}}{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \end{align} $$ (B1) 其中, $ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $如式(21)和式(23)所示.
利用引理1将式(65)的积分区域$ \mathbb{Z}_{0, {{m}^{1:s}}}^{1:s} $, $ \mathbb{Z}_{1, {{m}^{1:s}}}^{1:s}, $ $ \cdots, \mathbb{Z}_{\infty , {{m}^{1:s}}}^{1:s} $, $ {{\sigma }^{2}}\left( {{u}^{1:s}} \right) $又可写为
$$ \begin{align} &{{\sigma }^{2}}\left( {{u}^{1:s}} \right) = \sum\limits_{{{m}^{s}} = 0}^{\infty }{\cdots \sum\limits_{{{m}^{1}} = 0}^{\infty }{\sum\limits_{n = 0}^{\infty }{\frac{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}{{{m}^{1}}!\cdots {{m}^{s}}!n!}}}}\times\\&\quad\sum\limits_{\hat{n} = 0}^{\infty }{\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}{\int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{n}}}{q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)}}}}}\times\\&\quad{{e}^{2}}\left( {{{\pmb X}}_{n}}, {{{\hat{{\pmb X}}}}_{{\hat{n}}}} \right){\rm d}{{x}_{1:n}}{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \end{align} $$ (B2) 再根据假设1可得:
$$ \begin{align} {{\sigma }^{2}}\left( {{u}^{1:s}} \right) = \, &\sum\limits_{{{m}^{s}} = 0}^ {\infty }{\cdots \sum\limits_{{{m}^{1}} = 0}^{\infty } {\sum\limits_{n = 0}^{\infty }{\frac{{{\Omega }_{n, {{m}^{1:s}}}} \left( {{u}^{1:s}} \right)}{{{m}^{1}}!\cdots {{m}^{s}}!n!}}}} \times\\& \sum\limits_{\hat{n} = 0, n+\hat{n}>0}^{\infty }\sum\limits_ {{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}}\frac{1}{\max (\hat{n}, n)} \times\\&\int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{n}}}}}{q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)} \times\\& \begin{matrix} {}\end{matrix} \Bigg( \underset{\tau \in {{\Gamma }_{\max (\hat{n}, n)}}}{\mathop{\min }}\, \sum\limits_{i = 1}^{\min (\hat{n}, n)}\min \left( {{c}^{2}}, {{\left\| {{x}_{i}}- {{{\hat{x}}}_{{{\tau }_{i}}}} \right\|}^{2}} \right)+\\&{{c}^{2}}\text{abs}\left( \hat{n}-n \right) \Bigg) {\rm d}{{x}_{1:n}}{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \end{align} $$ (B3) 记
$$ \begin{align} {{\tau }^{*}} = \arg \underset{\tau \in {{\Gamma }_{\max (\hat{n}, n)}}}{\mathop{\min }}\, \left\{ \sum\limits_{i = 1}^{\min (\hat{n}, n)}{\min \left( {{c}^{2}}, {{\left\| {{x}_{i}}-{{{\hat{x}}}_{{{\tau }_{i}}}} \right\|}^{2}} \right)} \right\} \end{align} $$ (B4) 表示$ {{\Gamma }_{\max \left( \hat{n}, n \right)}} $中使得$ \sum\nolimits_{i = 1}^{\min \left( \hat{n}, n \right)}{\min \left( {{c}^{2}}, {{\left\| {{x}_{i}}-{{{\hat{x}}}_{{{\tau }_{i}}}} \right\|}^{2}} \right)} $最小的排序, 那么式(B3)又可写为
$$ \begin{align} & {{\sigma }^{2}}\left( {{u}^{1:s}} \right) = \\& \sum\limits_{{{m}^{s}} = 0}^{\infty }\cdots \sum\limits_{{{m}^{1}} = 0}^{\infty }\sum\limits_{n = 0}^{\infty } {\sum\limits_{\hat{n} = 0, n+\hat{n}>0}^{\infty }} \frac{{{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)} {{{m}^{1}}!\cdots {{m}^{s}}!n!}\cdot \\ &\frac{1}{\max (\hat{n}, n)}\Bigg( \sum\limits_{i = 1}^{\min (\hat{n}, n)}\min \Bigg( {{c}^{2}}, \frac{1}{{{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right)}\\&\sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}} \int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}{\cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}{\int_{{{\mathbb{X}}_{n}}} {q\left( \left. {{X}_{n}}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)}}} \times\\ &{{\left\| {{x}_{i}}-{{{\hat{x}}}_{\tau _{i}^{*}}} \right\|}^{2}}{\rm d}{{x}_{1:n}}{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \Bigg)+ {{c}^{2}}\text{abs}\left( \hat{n}-n \right) \Bigg) \end{align} $$ (B5) 其中, $ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $如式(22)和(24)所示.
利用式(29)定义的边缘密度$ {{q}_{n}}\left( \left. {\pmb x}, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $, 式(69)中的积分项可重写为
$$ \begin{align} & \sum\limits_{{{\ell }_{1:n}}\in {{\mathbb{L}}_{n}}} \int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}\cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}\int_{{{\mathbb{X}}_{n}}} q\left( \left. {{{\pmb X}}_{n}}, Z_{{{m}^{1:s}}}^{1:s}\times \right| {{u}^{1:s}} \right)\times\\&\qquad {{\left\| {{x}_{i}}-{{{\hat{x}}}_ {\tau _{i}^{*}}} \right\|}^{2}} d{{x}_{1:n}}dz_{1:{{m}^{1}}}^{1}\cdots dz_{1:{{m}^{s}}}^{s} = \\ &\qquad \int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}} \cdots \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}} \int_{{{\mathbb{X}}_{1}}}{{q}_{n}} \left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)\times\\&\qquad \sum\limits_{l = 1}^{L}{{{\left( {{x}^{l}}-\hat{x}_ {\tau _{i}^{*}}^{l} \right)}^{2}}} {\rm d}x{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \end{align} $$ (B6) 式中, $ {{q}_{n}}\left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right) $为式(31)的无标签版.
根据假设3可知估计器为无偏的, 那么可对式(70)应用式(4)所示的RFS观测下的信息不等式,
$$ \begin{align} \int_{\mathbb{Z}_{\hat{n}, {{m}^{s}}}^{s}}\!\cdots& \int_{\mathbb{Z}_{\hat{n}, {{m}^{1}}}^{1}}\int_{{{\mathbb{X}}_{1}}} {{q}_{n}}\left( \left. x, Z_{{{m}^{1:s}}}^{1:s} \right|{{u}^{1:s}} \right)\times\\&{{\left( {{x}^{l}}-\hat{x}_{\tau _{i}^{*}}^{l} \right)}^{2}}{\rm d}x{\rm d}z_{1:{{m}^{1}}}^{1}\cdots {\rm d}z_{1:{{m}^{s}}}^{s} \ge \\&{{\left[ J_{\hat{n}, n, {{m}^{1:s}}}^{-1} \left( {{u}^{1:s}} \right) \right]}^{l, l}}, \quad l = 1, \cdots, L \end{align} $$ (B7) 其中, $ {{J}_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $如式(35)所示.
将式(71)代入式(70)后再将所得结果代入式(69), 最终可得式(32)~(34).
-
图 1 单次实验中CS散度+穷尽法和误差界+穷尽法的传感器运动轨迹(黑实线表示目标运动轨迹, $ \bigcirc $和$ \Delta $分别表示目标的起点和终点; 灰色实线表示传感器运动轨迹, $ \square $和其上方数字分别表示传感器位置和处于该位置的时刻, $ \blacksquare $和$ \boxplus $分别表示传感器的起点和终点)
Fig. 1 Sensor trajectories in a simulation by using CS divergence with exhaustive search method and error bound with exhaustive search method(The black line is the target trajectory, $ \bigcirc $ and $ \Delta $ are the target starting point and ending point; the gray line is the sensor trajectory, $ \square $ and the number above it are the sensor position and the time when the sensor is located at the position, $ \blacksquare $ and $ \boxplus $ are the sensor starting point and ending point)
表 1 本文多目标MSE界的计算步骤
Table 1 The calculation steps of the multi-target MSE bound in this paper
步骤 计算多目标MSE界 第1步: 根据式(17)所示的多目标$ \delta $-GLMB预测密度$ {{\pi }_+}({\pmb X}) $产生$ M $个采样集$ {{\Psi}_{S}} = \left\{\tilde{{\pmb X}}_+^{(1)}, \cdots, \tilde{{\pmb X}}_+^{(M)} \right\} $; 第2步: 对于每一个$ \tilde{{\pmb X}}_+^{(j)}\in {{\Psi }_{S}} $, 根据式(7)所示的似然函数$ {{g}^{i}}\left( \cdot \left| \tilde{{\pmb X}}_+^{(j)}, {{u}^{i}} \right. \right) $, $ i = 1, \cdots, s $, 产生s个传感器的PIMS$ \tilde{Z}_{{{m}^{1:s}}}^{1:s, (j)} $; 第3步: 根据式(18)和(19)将所有的PIMS$ \left\{ \tilde{Z}_{{{m}^{1:s}}}^{1:s, (j)} \right\}_{j = 1}^{M} $分别划分到各个测量子空间$ \mathbb{Z}_{0, {{m}^{1:s}}}^{1:s}, \mathbb{Z}_{1, {{m}^{1:s}}}^{1:s}, \cdots, \mathbb{Z}_{\infty , {{m}^{1:s}}}^{1:s} $; 第4步: 给定$ \mathbb{Z}_{\hat{n}, {{m}^{1:s}}}^{1:s} $所包含的PIMS, 对式(24)和(35)应用伪MC积分法可得$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{J}_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $; 第5步: 根据式(23)、(33)和(34)计算得到$ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $, $ {{\Phi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $和$ {{\varepsilon }_{\hat{n}, n}} $; 第6步: 将所求得的$ {{\varpi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $, $ {{\Omega }_{n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $, $ {{\varepsilon }_{\hat{n}, n}} $和$ {{\Phi }_{\hat{n}, n, {{m}^{1:s}}}}\left( {{u}^{1:s}} \right) $代入式(32)可得$ {{\underline{\sigma }}^{2}}\left( {{u}^{1:s}} \right) $. 表 2 混合罚函数法
Table 2 Mixed penalty function method
步骤 混合罚函数求解过程 第1步: 选定满足不等式约束的初始点$ u_{(0)}^{1:s} $, 初始障碍因子$ r $及其缩小系数$ 0<C<1 $, 令$ i = 0 $; 第2步: 从$ j = 1 $到$ j = s $依次求$ u_{(i + 1)}^j = \arg {\mkern 1mu} \;\mathop {\min }\limits_{{u^j}} F\left( {u_{(i + 1)}^1, \cdots ,u_{(i + 1)}^{j - 1},{u^j},u_{(i)}^{j + 1}, \cdots ,u_{(i)}^s,r} \right) $, 此处只有$ {{u}^{j}} $为变量其余均为常量; 第3步: 若$ \sum\nolimits_{j = 1}^{s}{\left\| u_{(i+1)}^{j}-u_{(i)}^{j} \right\|}\le {{\varepsilon }_{2}} $, 转第4步; 否则令$ i = i+1 $, 转第2步; 第4步: 若$ \sum\nolimits_{j=1}^{s}{\left\| u_{(i+1)}^{j}-u_{(0)}^{j} \right\|}\le {{\varepsilon }_{1}} $, 输出$ u_{(i+1)}^{1:s} $为式(15)的解; 否则令$ r=Cr $, $ u_{(0)}^{1:s}=u_{(i+1)}^{1:s} $, $ i=0 $, 转第2步. 表 3 复合形法
Table 3 Complex method
步骤 复合形法求解过程 第1步: 选定$ {{\vec{u}}_{1}}, \cdots, {{\vec{u}}_{k}}\in{{\vec{\mathbb{U}}}_{D}} $ $ (N+2\le k\le2N) $构成初始复合形的$ k $个顶点, 为了避免降维, 要求至少有$ N+1 $个点的连线所形成的矢量有$ N $个是线性独立的; 第2步: 找出$ k $个顶点中目标函数值最大的点(最坏点)$ {{\vec{u}}_{H}} $和最小的点(最好点) $ {{\vec{u}}_{L}} $, $ {{\underline{\sigma }}^{2}} \left( {{{\vec{u}}}_{H}} \right) = \max\left\{ {{{\underline{\sigma }}}^{2}} \left( {{{\vec{u}}}_{1}}\right), \cdots, {{{\underline{\sigma }}}^{2}} \left({{{\vec{u}}}_{k}} \right) \right\} $, $ {{\underline{\sigma}}^{2}}\left( {{{\vec{u}}}_{L}} \right) = \min \left\{{{{\underline{\sigma }}}^{2}}\left( {{{\vec{u}}}_{1}}\right), \cdots, {{{\underline{\sigma }}}^{2}}\left({{{\vec{u}}}_{k}}\right) \right\} $; 第3步: 求除去$ {{\vec{u}}_{H}} $外其余$ k-1 $个顶点的中心点$ {{\vec{u}}_{0}} = \frac{1}{k-1}\left( \sum\limits_{i = 1}^{k}{{{{\vec{u}}}_{i}}}-{{{\vec{u}}}_{H}} \right) $.若$ {{\vec{u}}_{0}}\in{{\vec{\mathbb{U}}}_{D}} $, 转第4步; 否则, 转第1步, 重新选择初始顶点; 第4步: 以$ {{\vec{u}}_{0}} $为轴心求$ {{\vec{u}}_{H}} $的映射点$ {{\vec{u}}_{R}} = {{\vec{u}}_{0}}+t\left({{{\vec{u}}}_{0}}- {{{\vec{u}}}_{H}} \right) $, 其中$ t>0 $称为映射系数, 一般先取$ t = 1.3 $.若$ {{\vec{u}}_{R}}\in{{\vec{\mathbb{U}}}_{D}} $, 转第5步; 否则, 令$ t = 0.5t $, 转第4步; 第5步: 若$ {{\underline{\sigma }}^{2}}\left({{{\vec{u}}}_{R}}\right)<{{\underline{\sigma }}^{2}}\left( {{{\vec{u}}}_{H}} \right) $, 令$ {{\vec{u}}_{H}} = {{\vec{u}}_{R}} $, 转第6步; 否则, 令$ t = 0.5t $,转第4步.若$ t $已很小(如小于$ {{10}^{-5}} $)还不能满足上式, 说明由$ {{\vec{u}}_{H}} $和$ {{\vec{u}}_{0}} $形成的映射方向不利.为改变映射方向, 令$ {{\vec{u}}_{H}} = {{\vec{u}}_{SH}} $, 其中$ {{\vec{u}}_{SH}} $为次坏点, 转第3步; 第6步: 若$ {{\left( \frac{1}{k}\sum\limits_{i=1}^{k}{{{\left[ {{{\underline{\sigma }}}^{2}}\left( {{{\vec{u}}}_{0}}\right)-{{{\underline{\sigma }}}^{2}}\left( {{{\vec{u}}}_{i}}\right)\right]}^{2}}} \right)}^{\frac{1}{2}}}\le {{\varepsilon }_{1}} $或$ \mathop {\max }\limits_{1 \le i \le k} {\mkern 1mu} \left\| {{{\vec u}_i} - {{\vec u}_0}} \right\| \le {\varepsilon _2} $, 输出$ {{\vec{u}}_{L}} $为式(15)的解; 否则, 转第2步. 表 4 本文多传感器控制及多目标跟踪算法的执行流程
Table 4 The implementation process of multi-sensor control and multi-target tracking algorithm proposed in this paper
第1步: 对于式(15)所示的优化问题, 采用穷尽搜索、混合罚函数或复合形法三者之一计算多传感器最优控制量$ {{\left[ {{u}^{1:s}} \right]}^{*}} $; 第2步: 执行控制命令$ {{\left[ {{u}^{1:s}} \right]}^{*}} $后, 可获得真实测量集$ {{Z}^{1:s}} $; 第3步: 将$ {{Z}^{1:s}} $代入$ \delta $-GLMB滤波器的更新步, 可得多目标$ \delta $-GLMB后验密度$ \pi \left( \left.{\pmb X} \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) $; 第4步: 根据假设3可得多目标状态估计集$ {{\hat{{\pmb X}}}_{{\hat{n}}}} $; 第5步: 将$ \pi \left( \left. {\pmb X} \right|{{Z}^{1:s}}, {{u}^{1:s}} \right) $代入$ \delta $-GLMB滤波器的预测步, 可得下一时刻的多目标$ \delta $-GLMB预测密度$ {{\pi }_+}({\pmb X}) $.转第1步. 表 5 本文多传感器控制算法与基于GCI的多传感器控制算法的比较
Table 5 A comparison between multi-sensor control algorithm proposed in this paper and multi-sensor control algorithm based on GCI
本文多传感器控制算法 基于GCI的多传感器控制算法 目标函数 多目标MSE界 CS散度或PEECS 滤波器 $ \delta $-GLMB 边缘$ \delta $-GLMB或LMB (两者均为$ \delta $-GLMB的近似) 融合结构 集中式 分布式 通讯量 需传递各个传感器的测量集 需传递各个传感器的局部后验密度 计算方式 并行处理 并行处理 表 6 500次MC仿真平均后的OSPA误差终值(m)和CPU耗时(s)
Table 6 500 MC run averages of final OSPA error distance (m) and CPU processing time (s)
传感器控制算法 s = 4, E = 100J s = 4, E = 100J s = 4, E = 100J s = 4, E = 100J s = 4, E = 100J CS散度+穷尽法 19.5m/22s 13.9m/153s 9.2m/1062s 5.8m/7259s 4.6m/50973s 误差界+穷尽法 8.7m/23s 6.0m/162s 4.6m/1128s 3.9m/7698s 3.6m/53168s 误差界+混合罚函数法 8.9m/20s 6.3m/91s 5.1m/204s 4.5m/365s 4.1m/578s 误差界+复合形法 9.0m/16s 6.4m/81s 5.1m/206s 4.4m/573s 4.0m/1296s 建议的多目标MSE界 8.0m 5.4m 4.1m 3.5m 3.2m -
[1] Bar-Shalom Y, Li X R. Multitarget-Multisensor Tracking: Principles and Techniques. Storrs, CT: YBS Publishing, 1995. [2] Blackman S, Popoli R. Design and Analysis of Modern Tracking Systems. Norwood, MA: Artech House, 1999. [3] Tharmarasa R, Kirubarajan T, Hernandez M L, Sinha A. PCRLB-based multisensor array management for multitarget tracking. IEEE Transactions on Aerospace and Electronic Systems, 2007, 43(2): 539-555 doi: 10.1109/TAES.2007.4285352 [4] Tharmarasa R, Kirubarajan T, Sinha A, Lang T. Decentralized sensor selection for large-scale multisensor-multitarget tracking. IEEE Transactions on Aerospace and Electronic Systems, 2011, 47(2): 1307-1324 doi: 10.1109/TAES.2011.5751260 [5] Mohammadi A, Asif A. Decentralized computation of the conditional posterior Cramé r-Rao lower bound: application to adaptive sensor selection, In: Proceedings of the 38th IEEE International Conference on Acoustics, Speech, and Signal Processing, Vancouver, British Columbia, Canada: IEEE, 2013. 5278-5282 [6] Mahler R. Advances in Statistical Multisource-Multitarget Information Fusion. Norwood, MA: Artech House, 2014. [7] Vo B T, Vo B N. Labeled random finite sets and multi-object conjugate priors. IEEE Transactions on Signal Processing, 2013, 61(13): 3460-3475 doi: 10.1109/TSP.2013.2259822 [8] Vo B N, Vo B T, Phung D. Labeled random finite sets and the bayes multi-target tracking filter. IEEE Transactions on Signal Processing, 2014, 62(24): 6554-6567 doi: 10.1109/TSP.2014.2364014 [9] Reuter S, Vo B T, Vo B N, Dietmayer K. The labeled multi-bernoulli filter. IEEE Transactions on Signal Processing, 2014, 62(12): 3246-3260 doi: 10.1109/TSP.2014.2323064 [10] Papi F, Vo B N, Vo B T, Fantacci C, Beard M. Generalized labeled multi-Bernoulli approximation of multi-object densities. IEEE Transactions on Signal Processing, 2015, 63(20): 5487-5497 doi: 10.1109/TSP.2015.2454478 [11] Castanón D A, Carin L. Stochastic control theory for sensor management. In Hero Ⅲ A O, Castanón D A, Cochran D, Kastella K (Eds.), Foundations and applications of sensor management (pp. 7-32). Springler (Chapter 2), 2008. [12] Mahler R, Zajic T. Probabilistic objective functions for sensor management, In: Proceedings of the SPIE, vol.5429 Signal Processing, Sensor Fusion, and Target Recognition XⅢ, Orlando, FL, USA, April 2004, 233-244 [13] Hoang H G, Vo B T. Sensor management for multi-target tracking via multi-Bernoulli filtering, Automatica, 2014, 50(4): 1135-1142 doi: 10.1016/j.automatica.2014.02.007 [14] Gostar A K, Hoseinnezhad R, Bab-Hadiashar A. Multi-Bernoulli sensor control via minimization of expected estimation errors. IEEE Transactions on Aerospace and Electronic Systems, 2015, 51(3): 1762-1773 doi: 10.1109/TAES.2015.140211 [15] Ristic B, Vo B N. Sensor control for multi-object state-space estimation using random finite sets. Automatica, 2010, 46(11): 1812-1818 doi: 10.1016/j.automatica.2010.06.045 [16] Ristic B, Vo B N, Clark D. A note on the reward function for phd filters with sensor control. IEEE Transactions on Aerospace and Electronic Systems, 2011, 47(2): 1521-1529 doi: 10.1109/TAES.2011.5751278 [17] Beard M, Vo B T, Vo B N, Arulampalam S. Sensor control for multi-target tracking using cauchy-schwarz divergence, In: Proceedings of the 18th International Conference on Information Fusion, Washington D.C., USA: 2015, 937-944 [18] Gostar A K, Hoseinnezhad R, Rathnayake T, Wang X, Bab-Hadiashar A. Constrained sensor control for labeled multi-Bernoulli filter using Cauchy-Schwarz divergence. IEEE Signal Processing Letters, 2017, 24(9): 1313-1317 doi: 10.1109/LSP.2017.2723924 [19] Jiang M, Yi W, Kong L. Multi-sensor control for multi-target tracking using Cauchy-Schwarz divergence, In: Proceedings of the 19th International Conference on Information Fusion, Heidelberg, Germany: 2016, 2059-2066 [20] Wang X, Hoseinnezhad R, Gostar A K, Rathnayake T, Xu B, Bab-Hadiashar A. Multi-sensor control for multi-object Bayes filters. Signal Processing, 2018, 142(1): 260-270 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=21443d37af419592fe4d7fda0adb0442 [21] Mahler R. Optimal/robust distributed data fusion: a unified approach, In: Proceedings of the SPIE, vol. 4052 Signal Processing, Sensor Fusion, and Target Recognition IX, Orlando, FL, USA: 2000, 128-138 [22] Fantacci C, Vo B N, Vo B T, Battistelli G, Chisci L. Consensus labeled random finite set filtering for distributed multi-object tracking, arXiv preprint arXiv: 1501.01579, 2015. [23] Wright S J. Coordinate descent algorithms, Mathematical Programming, 2015, 151(1): 3-34 [24] Schuhmacher D, Vo B T, Vo B N. A consistent metric for performance evaluation of multi-object filters. IEEE Transactions on Signal Processing, 2008, 56(8): 3447-3457 doi: 10.1109/TSP.2008.920469 [25] Rezaeian M, Vo B N. Error bounds for joint detection and estimation of a single object with random finite set observation. IEEE Transactions on Signal Processing, 2010, 58(3): 1493-1506 doi: 10.1109/TSP.2009.2037665 [26] Zuo L, Niu R, Varshney P K. Conditional posterior Cramé r-Rao lower bounds for nonlinear sequential Bayesian estimation. IEEE Transactions on Signal Processing, 2011, 59(1): 1-14. http://dl.acm.org/citation.cfm?id=2199650 [27] Luenberger D G, Ye Y. Linear and Nonlinear Programming (4th Edition). New York: Springer, 2015. [28] Krus P, Andersson J. Optimizing optimization for design optimization, In: Proceedings of the ASME Design Automation Conference, Chicago, IL, USA: 2003, 951-960 [29] Poor V. An Introduction to Signal Detection and Estimation, New York: Springer-Verlag, 1994. https://www.springer.com/gp/book/9780387941738 [30] Davis P J, Rabinowitz P, Rheinbolt W. Methods of Numerical Integration (Second Edition). Mineola, New York: Dover Publications, 2007. [31] Mishra M, Gupta N, Dubey A, Shekhar S. Application of quasi Monte Carlo integration technique in efficient capacitance computation. Progress In Electromagnetics Research, 2009, 90: 309-322 doi: 10.2528/PIER09011310 [32] Krus P, Ölvander J. Performance index and meta optimization of a direct search optimization method. Engineering Optimization, 2013, 45(10): 1167-1185 doi: 10.1080/0305215X.2012.725052 [33] Li X R, Jilkov V P. Survey of maneuvering target tracking, Part V: multiple-model methods. IEEE Transactions on Aerospace and Electronic Systems, 2005, 41(4): 1255-1321 doi: 10.1109/TAES.2005.1561886 [34] Arulampalam S, Maskell S, Gordon N J, Clapp T. A tutorial on particle filters for on-line non-linear/non-Gaussian Bayesian tracking. IEEE Transactions on Signal Processing, 2002, 50(2): 174-188 doi: 10.1109/78.978374 [35] Poor V.An Introduction to Signal Detection and Estimation. New York: Springer-Verlag, 1994. 期刊类型引用(4)
1. 卢操,徐长飞,余洁文. 基于多目标的新能源微电网混合储能控制系统. 能源与环保. 2023(01): 267-272 . 百度学术
2. 何海祝,乔世成. 基于大数据的网络信息安全认证仿真研究. 计算机仿真. 2023(08): 398-402 . 百度学术
3. 张荆沙,张琰. 基于嵌入式技术的光纤干涉型传感器智能控制研究. 激光杂志. 2022(10): 193-196 . 百度学术
4. 张开利,李俊. 基于属性排查的软件运行风险多目标控制仿真. 计算机仿真. 2021(09): 306-309+385 . 百度学术
其他类型引用(6)
-