-
摘要: 高斯过程回归（Gaussian process regression，GPR）是一种广泛应用的回归方法，可以用于解决输入输出均为多元变量的人体姿态估计问题.计算复杂度是高斯过程回归的一个重要考虑因素，而常用的降低计算复杂度的方法为稀疏表示算法.在稀疏算法中，完全独立训练条件（Fully independent training conditional，FITC）法是一种较为先进的算法，多用于解决输入变量彼此之间完全独立的回归问题.另外，输入变量的噪声问题是高斯过程回归的另一个需要考虑的重要因素.对于测试的输入变量噪声，可以通过矩匹配的方法进行解决，而训练输入样本的噪声则可通过将其转换为输出噪声的方法进行解决，从而得到更高的计算精度.本文基于以上算法，提出一种基于噪声输入的稀疏高斯算法，同时将其应用于解决人体姿态估计问题.本文实验中的数据集来源于之前的众多研究人员，其输入为从视频序列中截取的图像或通过特征提取得到的图像信息，输出为三维的人体姿态.与其他算法相比，本文的算法在准确性，运行时间与算法稳定性方面均达到了令人满意的效果.Abstract: Gaussian process regression (GPR) is a common method for structured prediction and human pose estimation, in which input and output are both multivariate. Computational complexity is a significant consideration of GP regression and it can be reduced by sparse Gaussian algorithm. The fully independent training conditional (FITC) algorithm is a good method for sparse Gaussian process, and it can be applied to fully-independent input problems. Input noise is another significant consideration of GP regression. Moment matching can be used to solve trial input noise while training input noise can be modeled as output noise to achieve higher accuracy. On the basis of above algorithms, this study proposes a sparse Gaussian process with input noise for human pose estimation. A dataset from multiple people is used for experiments, in which the input is the image from video processing or image descriptor obtained by feature extraction, and the output is a three-dimensional human pose. The accuracy, runtime and stability of the algorithm are all satisfactory compared with other methods for human pose estimation.
-
近些年来, 由于多智能体协同控制在编队控制[1]、机器人网络[2]、群集行为[3]、移动传感器[4-5]等方面的广泛应用, 多智能体系统的协同控制问题受到了众多研究者的广泛关注.一致性问题是多智能体系统协同控制领域的一个关键问题, 其目的是通过与邻居之间的信息交换, 使所有智能体的状态达成一致.迄今为止, 对多智能体一致性的研究也已取得了丰硕的成果, 根据多智能体的动力学模型分类, 主要可以将其分为以下4种情形:一阶[6-9]、二阶[10-13]、三阶[14-15]、高阶[16-18].
在实际应用中, 由于CPU处理速度和内存容量的限制, 智能体不能频繁地进行控制以及与其邻居交换信息.因此, 事件触发控制策略作为减少控制次数和通信负载的有效途径, 受到了越来越多的关注.到目前为止, 对事件触发控制机制的研究也取得了很多成果[19-23].Xiao等[19]基于事件触发控制策略, 解决了带有领航者的离散多智能体系统的跟踪问题.通过利用状态测量误差并且基于二阶离散多智能体系统动力学模型, Zhu等[20]提出了一种自触发的控制策略, 该策略使得所有智能体的状态均达到一致. Huang等[21]研究了基于事件触发策略的Lur$'$e网络的跟踪问题.针对不同的领航者-跟随者系统, Xu等[22]提出了3种不同类型的事件触发控制器, 包含分簇式控制器、集中式控制器和分布式控制器, 以此来解决对应的一致性问题.然而, 大多数现有的事件触发一致性成果集中于考虑一阶多智能体系统和二阶多智能体系统, 很少有成果研究三阶多智能体系统的事件触发控制问题, 特别是对于三阶离散多智能体系统, 成果更是少之又少.所以, 设计相应的事件触发控制协议来解决三阶离散多智能体系统的一致性问题已变得尤为重要.
本文研究了基于事件触发控制机制的三阶离散多智能体系统的一致性问题, 文章主要有以下三点贡献:
1) 利用位置、速度和加速度三者的测量误差, 设计了一种新颖的事件触发控制机制.
2) 利用不等式技巧, 分析得到了保证智能体渐近收敛到一致状态的充分条件.与现有的事件触发文献[19-22]不同的是, 所得的一致性条件与通信拓扑的Laplacian矩阵特征值和系统的耦合强度有关.
3) 给出了排除类Zeno行为的参数条件, 进而使得事件触发控制器不会每个迭代时刻都更新.
1. 预备知识
1.1 代数图论
智能体间的通信拓扑结构用一个有向加权图来表示, 记为.其中, $\vartheta = \left\{ {1, 2, \cdots, n} \right\}$表示顶点集, $\varsigma\subseteq\vartheta\times\vartheta$表示边集, 称作邻接矩阵, ${a_{ij}}$表示边$\left({j, i} \right) \in \varsigma $的权值.当$\left({j, i} \right) \in \varsigma $时, 有${a_{ij}} > 0$; 否则, 有${a_{ij}} = 0$. ${a_{ij}} > 0$表示智能体$i$能收到来自智能体$j$的信息, 反之则不成立.对任意一条边$j$, 节点$j$称为父节点, 节点$i$则称为子节点, 节点$i$是节点$j$的邻居节点.假设通信拓扑中不存在自环, 即对任意$i\in \vartheta $, 有${a_{ii}} = 0$.
定义$L = \left({{l_{ij}}}\right)\in{\bf R}^{n\times n}$为图${\cal G}$的Laplacian矩阵, 其中元素满足${l_{ij}} = - {a_{ij}} \le 0, i \ne j$; ${l_{ii}} = \sum\nolimits_{j = 1, j \ne i}^n {{a_{ij}} \ge 0} $.智能体$i$的入度定义为${d_i} = \sum\nolimits_{j = 1}^n {{a_{ij}}} $, 因此可得到$L = D - \Delta $, 其中, .如果有向图中存在一个始于节点$i$, 止于节点$j$的形如的边序列, 那么称存在一条从$i$到$j$的有向路径.特别地, 如果图中存在一个根节点, 并且该节点到其他所有节点都有有向路径, 那么称此有向图存在一个有向生成树.另外, 如果有向图${\cal G}$存在一个有向生成树, 则Laplacian矩阵$L$有一个0特征值并且其他特征值均含有正实部.
1.2 模型描述
考虑多智能体系统由$n$个智能体组成, 其通信拓扑结构由有向加权图${\cal G}$表示, 其中每个智能体可看作图${\cal G}$中的一个节点, 每个智能体满足如下动力学方程:
$ \begin{equation} \left\{ \begin{array}{l} {x_i}\left( {k + 1} \right) = {x_i}\left( k \right) + {v_i}\left( k \right)\\ {v_i}\left( {k + 1} \right) = {v_i}\left( k \right) + {z_i}\left( k \right)\\ {z_i}\left( {k + 1} \right) = {z_i}\left( k \right) + {u_i}\left( k \right) \end{array} \right. \end{equation} $
(1) 其中, ${x_i}\left(k \right) \in \bf R$表示位置状态, ${v_i}\left(k \right) \in \bf R$表示速度状态, ${z_i}\left(k \right) \in \bf R$表示加速度状态, ${u_i}\left(k \right) \in \bf R$表示控制输入.
基于事件触发控制机制的控制器协议设计如下:
$ \begin{equation} {u_i}\left( k \right) = \lambda {b_i}\left( {k_p^i} \right) + \eta {c_i}\left( {k_p^i} \right) + \gamma {g_i}\left( {k_p^i} \right), k \in \left[ {k_p^i, k_{p + 1}^i} \right) \end{equation} $
(2) 其中, $\lambda> 0$, $\eta> 0$, $\gamma> 0$表示耦合强度,
$ \begin{align*}&{b_i}\left( k \right)= \sum\nolimits_{j \in {N_i}} {{a_{ij}}\left( {{x_j}\left( k \right) - {x_i}\left( k \right)} \right)} , \nonumber\\ &{c_i}\left( k \right)=\sum\nolimits_{j \in {N_i}} {{a_{ij}}\left( {{v_j}\left( k \right) - {v_i}\left( k \right)} \right)}, \nonumber\\ & {g_i}\left( k \right)=\sum\nolimits_{j \in {N_i}} {{a_{ij}}\left( {{z_j}\left( k \right) - {z_i}\left( k \right)} \right)} .\end{align*} $
触发时刻序列定义为:
$ \begin{equation} k_{p + 1}^i = \inf \left\{ {k:k > k_p^i, {E_i}\left( k \right) > 0} \right\} \end{equation} $
(3) ${E_i}\left(k \right)$为触发函数, 具有以下形式:
$ \begin{align} {E_i}\left( k \right)= & \left| {{e_{bi}}\left( k \right)} \right| + \left| {{e_{ci}}\left( k \right)} \right| + \left| {{e_{gi}}\left( k \right)} \right|- {\delta _2}{\beta ^k} - \nonumber\nonumber\\ &{\delta _1}\left| {{b_i}\left( {k_p^i} \right)} \right| - {\delta _1}\left| {{c_i}\left( {k_p^i} \right)} \right| - {\delta _1}\left| {{g_i}\left( {k_p^i} \right)} \right| \end{align} $
(4) 其中, ${\delta _1} > 0$, ${\delta _2} > 0$, $\beta > 0$, , ${e_{ci}}\left(k \right) = {c_i}\left({k_p^i} \right) - {c_i}\left(k \right)$, ${e_{gi}}\left(k \right) = {g_i}\left({k_p^i} \right) - {g_i}\left(k \right)$.
令$\varepsilon _i\left(k\right)={x_i}\left(k\right)-{x_1}\left(k\right)$, ${\varphi _i}\left(k\right)={v_i}\left(k \right)-$ ${v_1}\left(k\right)$, ${\phi _i}(k) = {z_i}(k) - {z_1}\left(k \right)$, $i = 2, \cdots, n$. , $\cdots, {\varphi _n}\left(k \right)]^{\rm T}$, $\phi \left(k \right) = {\left[{{\phi _2}\left(k \right), \cdots, {\phi _n}\left(k \right)} \right]^{\rm T}}$. $\psi \left(k \right) = {\left[{{\varepsilon ^{\rm T}}\left(k \right), {\varphi ^{\rm T}}\left(k \right), {\phi ^{\rm T}}\left(k \right)} \right]^{\rm T}}$, , ${\bar e_b} = {\left[{{e_{b1}}\left(k \right), \cdots, {e_{b1}}\left(k \right)} \right]^{\rm T}}$, , ${e_{c1}}\left(k \right)]^{\rm T}$, , ${\bar e_g} = $ ${\left[{{e_{g1}}\left(k \right), \cdots, {e_{g1}}\left(k \right)} \right]^{\rm T}}$, $\tilde e\left(k \right) = [\tilde e_b^{\rm T}\left(k \right), \tilde e_c^{\rm T}\left(k \right), $ $\tilde e_g^{\rm T}\left(k \right)]^{\rm T}$, $\bar e\left(k \right) = [\bar e_b^{\rm T}\left(k \right), \bar e_c^T\left(k \right), \bar e_g^{\rm T}\left(k \right)]^{\rm T}$,
$ \hat L = \left[ {\begin{array}{*{20}{c}} {{d_2} + {a_{12}}}&{{a_{13}} - {a_{23}}}& \cdots &{{a_{1n}} - {a_{2n}}}\\ {{a_{12}} - {a_{32}}}&{{d_3} + {a_{13}}}& \cdots &{{a_{1n}} - {a_{3n}}}\\ \vdots & \vdots & \ddots & \vdots \\ {{a_{12}} - {a_{n2}}}&{{a_{13}} - {a_{n3}}}& \cdots &{{d_n} + {a_{1n}}} \end{array}} \right] $
再结合式(1)和式(2)可得到:
$ \begin{equation} \psi \left( {k + 1} \right) = {Q_1}\psi \left( k \right) + {Q_2}\left( {\tilde e\left( k \right) - \bar e\left( k \right)} \right) \end{equation} $
(5) 其中, , .
定义1.对于三阶离散时间多智能体系统(1), 当且仅当所有智能体的位置变量、速度变量、加速度变量满足以下条件时, 称系统(1)能够达到一致.
$ \begin{align*} &{\lim _{k \to \infty }}\left\| {{x_j}\left( k \right) - {x_i}\left( k \right)} \right\| = 0 \nonumber\\ & {\lim _{k \to \infty }}\left\| {{v_j}\left( k \right) - {v_i}\left( k \right)} \right\| = 0 \nonumber\\ & {\lim _{k \to \infty }}\left\| {{z_j}\left( k \right) - {z_i}\left( k \right)} \right\| = 0 \\&\quad\qquad \forall i, j = 1, 2, \cdots , n \end{align*} $
定义2.如果$k_{p + 1}^i - k_p^i > 1$, 则称触发时刻序列$\left\{ {k_p^i} \right\}$不存在类Zeno行为.
假设1.假设有向图中存在一个有向生成树.
2. 一致性分析主要结果
假设$\kappa$是矩阵${Q_1}$的特征值, ${\mu _i}$是$L$的特征值, 则有如下等式成立:
$ {\rm{det}}\left( {\kappa {I_{3n - 3}} - {Q_1}} \right)=\nonumber\\ \det \left(\! \!{\begin{array}{*{20}{c}} {\left( {\kappa - 1} \right){I_{n - 1}}}\!&\!{ - {I_{n - 1}}}\!&\!{{0_{n - 1}}}\\ {{0_{n - 1}}}\!&\!{\left( {\kappa - 1} \right){I_{n - 1}}}\!&\!{ - {I_{n - 1}}}\\ {\lambda {{\hat L}_{n - 1}}}\!&\!{\eta {{\hat L}_{n - 1}}}\!&\!{\left( {\kappa - 1} \right){I_{n - 1}} + \gamma {{\hat L}_{n - 1}}} \end{array}} \!\!\right)=\nonumber\\ \prod\limits_{i = 2}^n {\left[ {{{\left( {\kappa - 1} \right)}^3} + \left( {\lambda + \eta \left( {\kappa - 1} \right) + \gamma {{\left( {\kappa - 1} \right)}^2}} \right){\mu _i}} \right]} $
令
$ \begin{align} {m_i}\left( \kappa \right)= &{\left( {\kappa - 1} \right)^3} + \nonumber\\&\left( {\lambda + \eta \left( {\kappa - 1} \right) + \gamma {{\left( {\kappa - 1} \right)}^2}} \right){\mu _i} = 0, \nonumber\\& \qquad\qquad\qquad\qquad\qquad i = 2, \cdots , n \end{align} $
(6) 则有如下引理:
引理1[15]. 如果矩阵$L$有一个0特征值且其他所有特征值均有正实部, 并且参数$\lambda $, $\eta $, $\gamma $满足下列条件:
$ \left\{ \begin{array}{l} 3\lambda - 2\eta < 0\\ \left( {\gamma - \eta + \lambda } \right)\left( {\lambda - \eta } \right) < - \dfrac{{\lambda \Re \left( {{\mu _i}} \right)}}{{{{\left| {{\mu _i}} \right|}^2}}}\\ \left( {4\gamma + \lambda - 2\eta } \right)<\dfrac{{8\Re \left( {{\mu _i}} \right)}}{{{{\left| {{\mu _i}} \right|}^2}}} \end{array} \right. $
那么, 方程(6)的所有根都在单位圆内, 这也就意味着矩阵${Q_1}$的谱半径小于1, 即$\rho \left({{Q_1}} \right) < 1$.其中, 表示特征值${\mu _i}$的实部.
引理2[23]. 如果, 那么存在$M \ge 1$和$0 < \alpha < 1$使得下式成立
$ {\left\| {{Q_1}} \right\|^k} \le M{\alpha ^k}, \quad k \ge 0 $
定理1. 对于三阶离散多智能体系统(1), 基于假设1, 如果式(2)中的耦合强度满足引理1中的条件, 触发函数(4)中的参数满足$0 < {\delta _1} < 1$, , $0 < \alpha < \beta < 1$, 则称系统(1)能够实现渐近一致.
证明.令$\omega \left(k \right) = \tilde e\left(k \right) - \bar e\left(k \right)$, 式(5)能够被重新写成如下形式:
$ \begin{equation} \psi \left( k \right) = Q_1^k\psi \left( 0 \right) + {Q_2}\sum\limits_{s = 0}^{k - 1} {Q_1^{k - 1 - s}\omega \left( s \right)} \end{equation} $
(7) 根据引理1和引理2可知, 存在$M \ge 1$和$0 < \alpha < 1$使得下式成立.
$ \begin{align} \left\| {\psi \left( k \right)} \right\|\le & {\left\| {{Q_1}} \right\|^k}\left\| {\psi \left( 0 \right)} \right\| + \nonumber\\ & \left\| {{Q_2}} \right\|\sum\limits_{s = 0}^{k - 1} {{{\left\| {{Q_1}} \right\|}^{k - 1 - s}}\left\| {\omega \left( s \right)} \right\|}\le \nonumber\\ & M\left\| {\psi \left( 0 \right)} \right\|{\alpha ^k}+\nonumber\\ & M\left\| {{Q_2}} \right\|\sum\limits_{s = 0}^{k - 1} {{\alpha ^{k - 1 - s}}\left\| {\omega \left( s \right)} \right\|} \end{align} $
(8) 由触发条件可得:
$ \begin{align} & \left| {{e_{bi}}\left( k \right)} \right| + \left| {{e_{ci}}\left( k \right)} \right| + \left| {{e_{gi}}\left( k \right)} \right|\le\nonumber\\ & \qquad{\delta _1}\left| {{b_i}\left( {k_p^i} \right)} \right| + {\delta _1}\left| {{c_i}\left( {k_p^i} \right)} \right| +\nonumber\\ &\qquad {\delta _1}\left| {{g_i}\left( {k_p^i} \right)} \right| + {\delta _2}{\beta ^k}\le\nonumber\\ &\qquad {\delta _1}\left\| L \right\| \cdot \left\| {\varepsilon \left( k \right)} \right\| + {\delta _1}\left\| L \right\| \cdot \left\| {\varphi \left( k \right)} \right\| + \nonumber\\ &\qquad{\delta _1}\left\| L \right\| \cdot \left\| {\phi \left( k \right)} \right\|+ {\delta _1}\left| {{e_{bi}} \left( k \right)} \right| + \nonumber\\ &\qquad{\delta _1}\left| {{e_{ci}} \left( k \right)} \right|+ {\delta _1}\left| {{e_{gi}}\left( k \right)} \right| + {\delta _2}{\beta ^k} \end{align} $
(9) 对上式移项可求解得:
$ \begin{align} &\left| {{e_{bi}}\left( k \right)} \right| + \left| {{e_{ci}}\left( k \right)} \right| + \left| {{e_{gi}}\left( k \right)} \right|\le \nonumber\\ &\qquad\frac{{{\delta _1}\left\| L \right\| \cdot \left\| {\varepsilon \left( k \right)} \right\|}}{{1 - {\delta _1}}} + \frac{{{\delta _1}\left\| L \right\| \cdot \left\| {\varphi \left( k \right)} \right\|}}{{1 - {\delta _1}}}{\rm{ + }}\nonumber\\ &\qquad\frac{{{\delta _1}}}{{1 - {\delta _1}}}\left\| L \right\| \cdot \left\| {\phi \left( k \right)} \right\| + \frac{{{\delta _2}}}{{1 - {\delta _1}}}{\beta ^k} \end{align} $
(10) 又因为, 和, 可得出下列不等式:
$ \begin{align} &\left| {{e_{bi}}\left( k \right)} \right| + \left| {{e_{ci}}\left( k \right)} \right| + \left| {{e_{gi}}\left( k \right)} \right|\le\nonumber\\ &\qquad \frac{{{\delta _1}\left\| L \right\|}}{{1 - {\delta _1}}} \cdot \left( {\left\| {\varepsilon \left( k \right)} \right\|{\rm{ + }}\left\| {\varphi \left( k \right)} \right\|{\rm{ + }}\left\| {\phi \left( k \right)} \right\|} \right) +\nonumber\\ &\qquad \frac{{{\delta _2}{\beta ^k}}}{{1 - {\delta _1}}}\le \frac{{3{\delta _1}}}{{1 - {\delta _1}}}\left\| L \right\| \cdot \left\| {\psi \left( k \right)} \right\| + \frac{{{\delta _2}}}{{1 - {\delta _1}}}{\beta ^k} \end{align} $
(11) 接着有如下不等式成立:
$ \begin{align} \left\| {e\left( k \right)} \right\|\le \frac{{3\sqrt n {\delta _1}}}{{1 - {\delta _1}}}\left\| L \right\| \cdot \left\| {\psi \left( k \right)} \right\| + \frac{{\sqrt n {\delta _2}}}{{1 - {\delta _1}}}{\beta ^k} \end{align} $
(12) 其中, , ${e_b}(k) = \left[{{e_{b1}}(k), \cdots, {e_{bn}}(k)} \right]$, ${e_c}(k) = \left[{{e_{c1}}(k), \cdots, {e_{cn}}(k)} \right]$,
注意到
$ \begin{equation} \left\| {\tilde e( k )} \right\| + \left\| {\bar e( k )} \right\| \le \sqrt {6( {n - 1} )} \left\| {e( k )} \right\| \end{equation} $
(13) 于是有
$ \begin{align} \left\| {\omega ( k )} \right\| &= \left\| {\tilde e( k ) - \bar e\left( k \right)} \right\| \le\nonumber\\ & \left\| {\tilde e\left( k \right)} \right\| + \left\| {\bar e\left( k \right)} \right\|\le\nonumber\\ & \frac{{3\sqrt {6n( {n - 1} )} {\delta _1}}}{{1 - {\delta _1}}}\left\| L \right\| \cdot \left\| {\psi \left( k \right)} \right\| +\nonumber\\ & \frac{{\sqrt {6n( {n - 1} )} {\delta _2}}}{{1 - {\delta _1}}}{\beta ^k} \end{align} $
(14) 把式(14)代入式(8)可得
$ \begin{align} \left\| {\psi \left( k \right)} \right\| &\le M\left\| {\psi \left( 0 \right)} \right\|{\alpha ^k}+ \nonumber\\ &\frac{{M\left\| {{Q_2}} \right\|{\alpha ^{k - 1}} {\delta _1}3\sqrt {6n\left( {n - 1} \right)} \left\| L \right\|}}{{1 - {\delta _1}}}\times\nonumber\\ &\sum\limits_{s = 0}^{k - 1} {{\alpha ^{ - s}}\left\| {\psi \left( s \right)} \right\|} + M\left\| {{Q_2}} \right\|{\alpha ^{k - 1}}\times\nonumber\\ &\sum\limits_{s = 0}^{k - 1} {{\alpha ^{ - s}} \frac{{\sqrt {6n\left( {n - 1} \right)} {\delta _2}}} {{1 - {\delta _1}}}{\beta ^s}} \end{align} $
(15) 接下来的部分, 将证明下列不等式成立.
$ \begin{equation} \left\| {\psi \left( k \right)} \right\| \le W{\beta ^k}.\end{equation} $
(16) 其中, $W = \max \left\{ {{\Theta _1}, {\Theta _2}} \right\}$,
首先, 证明对任意的$\rho > 1$, 下列不等式成立.
$ \begin{equation} \left\| {\psi \left( k \right)} \right\| < \rho W{\beta ^k} \end{equation} $
(17) 利用反证法, 先假设式(17)不成立, 则必将存在${k^ * } > 0$使得并且当$k \in \left({0, {k^ * }} \right)$时$\left\| {\psi \left(k \right)} \right\| < \rho W{\beta ^k}$成立.因此, 根据式(17)可得:
$ \begin{align*} &\rho W{\beta ^{{k^ * }}} \le \left\| {\psi \left( {{k^ * }} \right)} \right\| \le\\ &\qquad M\left\| {\psi \left( 0 \right)} \right\|{\alpha ^{{k^ * }}} +\left\| {{Q_2}} \right\|{\alpha ^{{k^ * } - 1}}M\times \end{align*} $
$ \begin{align*} &\qquad\sum\limits_{s = 0}^{{k^ * } - 1} {\alpha ^{ - s}}\left[ {\frac{{3\sqrt {6n\left( {n - 1} \right)} {\delta _1}\left\| L \right\| \cdot \left\| {\psi \left( s \right)} \right\|}}{{1 - {\delta _1}}}} \right]+ \\ &\qquad M\left\| {{Q_2}} \right\|{\alpha ^{{k^ * } - 1}} \sum\limits_{s = 0}^{{k^ * } - 1} {{\alpha ^{ - s}} \left[ {\frac{{\sqrt {6n\left( {n - 1} \right)} {\delta _2}}}{{1 - {\delta _1}}}{\beta ^s}} \right]} < \\ &\qquad \rho M\left\| {\psi \left( 0 \right)} \right\|{\alpha ^{{k^ * }}} + \rho M\left\| {{Q_2}} \right\|{\alpha ^{{k^ * } - 1}}\times\\ &\qquad \sum\limits_{s = 0}^{{k^ * } - 1} {{\alpha ^{ - s}} \left[ {\frac{{3\sqrt {6n\left( {n - 1} \right)} {\delta _1}\left\| L \right\| \cdot W{\beta ^s}}} {{1 - {\delta _1}}}} \right]} +\\ &\qquad\rho M\left\| {{Q_2}} \right\|{\alpha ^{{k^ * } - 1}} \sum\limits_{s = 0}^{{k^ * } - 1} {{\alpha ^{ - s}} \left[ {\frac{{\sqrt {6n\left( {n - 1} \right)} {\delta _2}{\beta ^s}}}{{1 - {\delta _1}}}} \right]=} \\ &\qquad \rho M\left\| {\psi \left( 0 \right)} \right\|{\alpha ^{{k^ * }}}- \nonumber\\ &\qquad \rho \frac{{M\left\| {{Q_2}} \right\|\sqrt {6n\left( {n - 1} \right)} \left( {3{\delta _1}\left\| L \right\|W + {\delta _2}} \right)}}{{\left( {\beta - \alpha } \right)\left( {1 - {\delta _1}} \right)}}{\alpha ^{{k^ * }}}+\nonumber\\ &\qquad \rho \frac{{M\left\| {{Q_2}} \right\|\sqrt {6n\left( {n - 1} \right)} \left( {3{\delta _1}\left\| L \right\|W + {\delta _2}} \right)}}{{\left( {\beta - \alpha } \right)\left( {1 - {\delta _1}} \right)}}{\beta ^{{k^ * }}} \end{align*} $
1) 当$W = M\left\| {\psi \left(0 \right)} \right\|$时, 则有
$ \begin{equation*} \begin{aligned} &M\left\| {\psi \left( 0 \right)} \right\| - \nonumber\\ &\qquad \frac{{M\left\| {{Q_2}} \right\|\sqrt {6n\left( {n - 1} \right)} \left( {3{\delta _1}\left\| L \right\|W + {\delta _2}} \right)}}{{\left( {\beta - \alpha } \right)\left( {1 - {\delta _1}} \right)}} \ge 0 \end{aligned} \end{equation*} $
所以可得到
$ \begin{equation} \rho W{\beta ^{{k^ * }}} \le \left\| {\psi \left( {{k^ * }} \right)} \right\| \le \rho M\left\| {\psi \left( 0 \right)} \right\|{\beta ^{{k^ * }}}=\rho W{\beta ^{{k^ * }}} \end{equation} $
(18) 2) 当时, 则有
$ \begin{equation*} \begin{aligned} &M\left\| {\psi \left( 0 \right)} \right\|- \nonumber\\ &\qquad\frac{{M\left\| {{Q_2}} \right\|\sqrt {6n\left( {n - 1} \right)} \left( {3{\delta _1}\left\| L \right\|W + {\delta _2}} \right)}}{{\left( {\beta - \alpha } \right)\left( {1 - {\delta _1}} \right)}} < 0 \end{aligned} \end{equation*} $
所以有
$ \begin{align} &\rho W{\beta ^{{k^ * }}} \le \left\| {\psi \left( {{k^ * }} \right)} \right\|\le\nonumber\\ & \frac{{\rho {\delta _2}M\left\| {{Q_2}} \right\|\sqrt {6n\left( {n - 1} \right)} {\beta ^{{k^ * }}}}}{{\left( {\beta - \alpha } \right)\left( {1 - {\delta _1}} \right) - 3{\delta _1}M\left\| {{Q_2}} \right\|\left\| L \right\|\sqrt {6n\left( {n - 1} \right)} }}=\nonumber\\ &\rho W{\beta ^{{k^ * }}} \end{align} $
(19) 根据以上结果, 式(18)和式(19)都与假设相矛盾.这说明原命题成立, 即对任意的$\rho > 1$, 式(17)成立.易知, 如果$\rho \to 1$, 则式(16)成立.根据式(16)可知, 当$k \to + \infty $时, 有, 则系统(5)是收敛的.由$\psi \left(k \right)$的定义可知, 系统(1)能够实现渐近一致.
定理2. 对于系统(1), 如果定理1中的条件成立, 并且控制器(2)中的设计参数满足如下条件,
$ {\delta _1} \in \left( {\frac{{\left( {\beta - \alpha } \right)}}{{\left( {\beta - \alpha } \right) + 3\sqrt {6n\left( {n - 1} \right)} M\left\| {{Q_{\rm{2}}}} \right\|\left\| L \right\|}}, 1} \right)\\ {\delta _2} > \frac{{\left\| L \right\|\left\| {\psi \left( 0 \right)} \right\|M\left( {1 + \beta } \right)}}{\beta } $
那么触发序列中的类Zeno行为将被排除.
证明. 易知排除类Zeno行为的关键是要证明不等式$k_{p + 1}^i - k_p^i > 1$成立.根据事件触发机制可知, 下一个触发时刻将会发生在触发函数(4)大于0时.进而可得到如下不等式
$ \begin{align} &\left| {{e_{bi}}\left( {k_{p + 1}^i} \right)} \right| + \left| {{e_{ci}}\left( {k_{p + 1}^i} \right)} \right| + \left| {{e_{gi}}\left( {k_{p + 1}^i} \right)} \right|\ge\nonumber\\ &\qquad{\delta _1}\left| {{b_i}\left( {k_p^i} \right)} \right| + {\delta _1}\left| {{c_i}\left( {k_p^i} \right)} \right| +\nonumber\\ &\qquad {\delta _1}\left| {{g_i}\left( {k_p^i} \right)} \right| + {\delta _2}{\beta ^{k_{p + 1}^i}} \end{align} $
(20) 定义, .结合式(20), 可得到下式
$ \begin{equation} {G_i}\left( {k_{p + 1}^i} \right) \ge {\delta _1}{H_i}\left( {k_p^i} \right) + {\delta _2}{\beta ^{k_{p + 1}^i}} \end{equation} $
(21) 结合式(16)和式(21)可得
$ \begin{align} {\delta _2}{\beta ^{k_{p + 1}^i}} &\le {G_i}\left( {k_{p + 1}^i} \right) - {\delta _1}{H_i}\left( {k_p^i} \right)\le\nonumber\\ & \left\| L \right\|\left( {\left\| {\psi \left( {k_p^i} \right)} \right\| + \left\| {\psi \left( {k_{p + 1}^i} \right)} \right\|} \right)\le\nonumber\\ & W\left\| L \right\|\left( {{\beta ^{k_p^i}} + {\beta ^{k_{p + 1}^i}}} \right) \end{align} $
(22) 求解上式得
$ \begin{equation} \left( {{\delta _2} - \left\| L \right\|W} \right){\beta ^{k_{p + 1}^i}} \le \left\| L \right\|W{\beta ^{k_p^i}} \end{equation} $
(23) 根据式(23)可得
$ \begin{equation} k_{p + 1}^i - k_p^i > \dfrac{{\ln \dfrac{{W\left\| L \right\|}}{{{\delta _2} - W\left\| L \right\|}}} } {\ln \beta } \end{equation} $
(24) 基于(24)易知当时, 有如下不等式成立
$ \begin{equation} \dfrac{{\ln \dfrac{{W\left\| L \right\|}}{{{\delta _2} - W\left\| L \right\|}}}} {\ln \beta } > 1 \end{equation} $
(25) 此外, 因为$W = M\left\| {\psi \left(0 \right)} \right\|$以及
$ \begin{equation} {\delta _1} > \frac{{\left( {\beta - \alpha } \right)}}{{\left( {\beta - \alpha } \right) + 3\sqrt {6n\left( {n - 1} \right)} M\left\| {{Q_{\rm{2}}}} \right\|\left\| L \right\|}} \end{equation} $
(26) 又可以得出
$ \begin{equation} {\delta _2} > \frac{{\left\| L \right\|\left\| {\psi \left( 0 \right)} \right\|M\left( {1 + \beta } \right)}}{\beta } = \frac{{\left\| L \right\|W\left( {1 + \beta } \right)}}{\beta } \end{equation} $
(27) 该式意味着式(25)成立, 又结合式(24)易知$k_{p + 1}^i - k_p^i > 1$, 即排除类Zeno行为的条件得已满足.
注2.类Zeno行为广泛存在于基于事件触发控制机制的离散系统中.然而, 当前极少有文献研究如何排除类Zeno行为, 尤其是对于三阶多智能体动态模型.定理2给出了排除三阶离散多智能体系统的类Zeno行为的参数条件.
3. 仿真实验
本部分将利用一个仿真实验来验证本文所提算法及理论的正确性和有效性.假设三阶离散多智能体系统(1)包含6个智能体, 且有向加权通信拓扑结构如图 1所示, 权重取值为0或1, 可以明显地看出该图包含有向生成树(满足假设1).
通过简单的计算可得, ${\mu _1} = 0$, ${\mu _2} = 0.6852$, ${\mu _3} = 1.5825 + 0.3865$i, ${\mu _4} = 1.5825 - 0.3865$i, ${\mu _5} = 3.2138$, ${\mu _6} = 3.9360$.令$M = 1$, 结合定理1和定理2可得到$0.035 < {\delta _1} < 1$, ${\delta _2} > 44.0025$, $0 < \alpha < \beta < 1$.令${\delta _1} = 0.2$, ${\delta _2} = 200$, $\alpha = 0.6$, $\beta = 0.9$, $\lambda = 0.02$, $\eta = 0.3$, $\gamma = 0.5$, 不难验证满足引理1的条件并且计算可知$\rho \left({{Q_1}} \right) = 0.9958 < 1$.三阶离散多智能体系统(1)的一致性结果如图 2~图 6所示.根据定理1可知, 基于控制器(2)和事件触发函数(4)的系统(1)能实现一致.从图 2~图 6可以看出, 仿真结果与理论分析符合.
图 2 三阶离散多智能体系统的位置轨迹图Fig. 2 The trajectories of position in third-order discrete-time multi-agent systems
图 3 三阶离散多智能体系统的速度轨迹图Fig. 3 The trajectories of speed in third-order discrete-time multi-agent systems
图 4 三阶离散多智能体系统的加速度轨迹图Fig. 4 The trajectories of acceleration in third-order discrete-time multi-agent systems
图 5 三阶离散多智能体系统的控制轨迹图Fig. 5 The trajectories of control in third-order discrete-time multi-agent systems图 2~图 4分别表征了系统(1)中所有智能体的位置、速度和加速度的轨迹, 从图中可以看出以上3个变量确实达到了一致.图 5展示了控制输入的轨迹.为了更清楚地体现事件触发机制的优点, 图 6给出了0$ \sim $100次迭代内的各智能体的触发时刻轨迹.从图 6可以看出, 本文设计的事件触发协议确实达到了减少更新次数, 节省资源的目的.
4. 结论
针对三阶离散多智能体系统的一致性问题, 构造了一个新颖的事件触发一致性协议, 分析得到了在通信拓扑为有向加权图且包含生成树的条件下, 系统中所有智能体的位置状态、速度状态和加速度状态渐近收敛到一致状态的充分条件.同时, 该条件指出了通信拓扑的Laplacian矩阵特征值和系统的耦合强度对系统一致性的影响.另外, 给出了排除类Zeno行为的参数条件.仿真实验结果也验证了上述结论的正确性.将文中获得的结论扩展到拓扑结构随时间变化的更高阶多智能体网络是极有意义的.这将是未来研究的一个具有挑战性的课题.
-
图 3 GP, KTA, HSICKNN与SGPIN算法的误差比较
Fig. 3 Error comparison of GP, KTA, HSICKNN and SGPIN
表 1 GP, FITC, NIGP和SGPIN算法比较
Table 1 Comparison of GP, FITC, NIGP and SGPIN
算法 训练点个数 MSE ($10^{-3}$) 运行时间(s) GP 200 31.1326 1.876034 FITC 800 18.6279 0.062001 NIGP 200 18.6279 13.630882 SGPIN 800 8.6265 0.003087 200 18.4946 0.002612 
下载: 导出CSV
表 2 实验数据集
Table 2 Experimental set
特征 动作 个体1 个体2 个体3 总数 HoG Walking 1 176 876 895 2 947 Jogging 439 795 831 2 065 Throw/Catch 217 806 0 1 023 Gestures 801 681 214 1 696 Box 502 464 933 1 889 Total 3 135 3 622 2 873 9 630
下载: 导出CSV
表 3 基于HumanEva-I数据集HoG特征的不同算法的平均误差
Table 3 Evaluation of average error of difierent algorithms based on HoG feature of HumanEva-I
研究个体 动作 样本数 GP TGP TGPKNN KTA HSICKNN SGPIN S1 Walking 1 176 398.5823 197.1179 193.9949 213.5265 218.6241 161.2112 Jogging 439 383.7747 212.3234 212.2018 188.6683 196.0839 154.5919 Throw/Catch 217 414.5873 174.2834 / / / 100.7592 Gestures 801 415.3106 98.6237 102.5520 92.1541 156.6464 20.1770 Box 502 426.6358 162.6801 163.3203 118.0500 149.5003 82.3949 S2 Walking 876 398.5817 197.1496 195.5694 206.7040 211.9735 160.4342 Jogging 795 405.1201 213.0572 207.2430 227.3562 231.1777 176.1768 Throw/Catch 806 421.5898 210.1543 199.3265 173.2717 189.7417 92.6742 Gestures 681 410.0671 201.1053 201.7576 153.9103 173.0548 63.2473 Box 464 421.3947 171.6007 109.1912 137.1031 159.5833 98.3920 S3 Walking 895 412.0019 219.2579 214.8589 236.1566 239.6487 177.3461 Jogging 831 441.7053 211.1343 206.1400 233.5746 236.5287 184.2251 Throw/Catch 0 / / / / / / Gestures 214 473.7616 159.7482 / / / 40.3100 Box 933 483.6534 214.1621 207.7578 186.5170 195.9815 120.6541 总数 9 630 284.0985 160.1196 162.0768 / / 155.3066
下载: 导出CSV
表 4 基于HumanEva-I数据集HoG特征的不同算法的运行时间
Table 4 Evaluation of runtime of difierent algorithms based on HoG feature of HumanEva-I
研究个体 动作 样本数 GP TGP TGPKNN KTA HSICKNN SGPIN S1 Walking 1 176 0.11 26.77 24.67 28.16 27.87 18.02 Jogging 439 0.03 8.47 10.43 10.18 10.26 21.65 Throw/Catch 217 0.01 3.77 / / / 22.44 Gestures 801 0.07 27.15 27.31 18.64 19.42 19.78 Box 502 0.03 10.11 11.19 11.75 11.84 21.90 S2 Walking 876 0.08 20.86 25.83 20.03 20.26 22.04 Jogging 795 0.07 18.06 17.86 17.64 17.74 23.32 Throw/Catch 806 0.02 18.56 26.59 20.13 20.02 21.69 Gestures 681 0.04 14.32 15.52 15.91 16.64 18.38 Box 464 0.03 9.02 10.35 10.71 11.43 23.69 S3 Walking 895 0.09 22.78 22.63 20.75 20.95 21.13 Jogging 831 0.08 21.83 20.13 18.51 19.01 20.36 Throw/Catch 0 / / / / / / Gestures 214 0.04 6.13 / / / 22.62 Box 933 0.10 23.70 23.67 22.68 23.57 21.69 总数 9 630 11 1928 442 491 495 41
下载: 导出CSV
表 5 个体3行走姿态的预测误差
Table 5 Predicting errors of subject 3 walking
GP TGP TGPKNN HSICKNN KTA SGPIN 1 412.0 219.1 214.8 236.3 239.6 177.6 2 412.0 218.0 214.9 232.5 235.7 176.9 3 412.0 220.2 214.6 237.1 240.9 177.7 4 412.0 220.4 215.6 241.6 244.8 177.5 5 412.0 218.7 214.5 233.4 237.3 177.0 方差 0.00 1.04 0.18 12.89 12.38 0.14
下载: 导出CSV
-
[1] 沈建冬, 陈恒.融合HOG和颜色特征的人体姿态估计新算法.计算机工程与应用, 2017, 53(21):190-194 doi: 10.3778/j.issn.1002-8331.1606-0319Shen Jian-Dong, Chen Heng. New human pose estimation algorithm based on HOG and color features. Computer Engineering and Applications, 2017, 53(21):190-194 doi: 10.3778/j.issn.1002-8331.1606-0319 [2] Wang J M, Fleet D J, Hertzmann A. Gaussian process dynamical models for human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008, 30(2):283-298 doi: 10.1109/TPAMI.2007.1167 [3] 袁紫华, 李峰, 周书仁.基于Haar型LBP纹理特征的人体姿态估计.计算机工程, 2015, 41(4):199-204 doi: 10.3969/j.issn.1000-3428.2015.04.038Yuan Zi-Hua, Li Feng, Zhou Shu-Ren. Human pose estimation based on Haar characteristics LBP texture feature. Computer Engineering, 2015, 41(4):199-204 doi: 10.3969/j.issn.1000-3428.2015.04.038 [4] Zhao X, Ning H Z, Liu Y C, Huang T. Discriminative estimation of 3D human pose using Gaussian processes. In:Proceedings of the 19th International Conference on Pattern Recognition. Tampa, FL, USA:IEEE, 2008. 1-4 [5] Bratieres S, Quadrianto N, Ghahramani Z. GPstruct:Bayesian structured prediction using gaussian processes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(7):1514-1520 doi: 10.1109/TPAMI.2014.2366151 [6] Ding M, Fan G L. Articulated Gaussian kernel correlation for human pose estimation. In:Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Boston, MA, USA:IEEE, 2015. 57-64 [7] Rasmussen C E, Ghahramani Z. Infinite mixtures of Gaussian process experts. In:Proceedings of the 14th International Conference on Neural Information Processing Systems:Natural and Synthetic. Vancouver, British Columbia, Canada:MIT Press, 2002. 881-888 [8] 俞斌峰, 季海波.稀疏贝叶斯混合专家模型及其在光谱数据标定中的应用.自动化学报, 2016, 42(4):566-579 http://www.aas.net.cn/CN/abstract/abstract18844.shtmlYu Bin-Feng, Ji Hai-Bo. Sparse Bayesian mixture of experts and its application to spectral multivariate calibration. Acta Automatica Sinica, 2016, 42(4):566-579 http://www.aas.net.cn/CN/abstract/abstract18844.shtml [9] 刘长红, 杨扬, 陈勇.增量式人体姿态映射模型的学习方法.计算机科学, 2010, 37(3):268-270 doi: 10.3969/j.issn.1002-137X.2010.03.067Liu Chang-Hong, Yang Yang, Chen Yong. Incrementally learning human pose mapping model. Computer Science, 2010, 37(3):268-270 doi: 10.3969/j.issn.1002-137X.2010.03.067 [10] 闫小喜, 韩崇昭.基于增量式有限混合模型的多目标状态极大似然估计.自动化学报, 2011, 37(5):577-584 http://www.aas.net.cn/CN/abstract/abstract17393.shtmlYan Xiao-Xi, Han Chong-Zhao. Maximum likelihood estimation of multiple target states based on incremental finite mixture model. Acta Automatica Sinica, 2011, 37(5):577-584 http://www.aas.net.cn/CN/abstract/abstract17393.shtml [11] Csató L, Opper M. Sparse on-line Gaussian processes. Neural Computation, 2002, 14(3):641-668 doi: 10.1162/089976602317250933 [12] Bijl H, van Wingerden J W, Schön T B, Verhaegen M. Online sparse Gaussian process regression using FITC and PITC approximations. IFAC-PapersOnLine, 2015, 48(28):703-708 doi: 10.1016/j.ifacol.2015.12.212 [13] Snelson E, Ghahramani Z. Sparse Gaussian processes using pseudo-inputs. In:Proceedings of the 18th International Conference on Neural Information Processing Systems. Vancouver, British Columbia, Canada:MIT Press, 2006. 1257-1264 [14] McHutchon A, Rasmussen C E. Gaussian process training with input noise. In:Proceedings of the 24th International Conference on Neural Information Processing Systems. Granada, Spain:ACM, 2011. 1341-1349 [15] HumanEva Dataset[Online], available:http://humaneva.is.tue.mpg.de/, November 3, 2017 [16] Sigal L, Balan A O, Black M J. HumanEva:synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 2006, 87(1-2):Article No. 4 http://d.old.wanfangdata.com.cn/NSTLQK/NSTL_QKJJ0224696659/ [17] Poppe R. Evaluating example-based pose estimation:experiments on the HumanEva sets. In:Proceedings of the 2007 Computer Vision and Pattern Recognition Workshop on Evaluation of Articulated Human Motion and Pose Estimation (EHuM2). Minneapolis, USA:IEEE, 2007. [18] 苏本跃, 蒋京, 汤庆丰, 盛敏.基于函数型数据分析方法的人体动态行为识别.自动化学报, 2017, 43(5):866-876 http://www.aas.net.cn/CN/abstract/abstract19064.shtmlSu Ben-Yue, Jiang Jing, Tang Qing-Feng, Sheng Min. Human dynamic action recognition based on functional data analysis. Acta Automatica Sinica, 2017, 43(5):866-876 http://www.aas.net.cn/CN/abstract/abstract19064.shtml [19] Shakhnarovich G, Viola P, Darrell T. Fast pose estimation with parameter-sensitive hashing. In:Proceedings of the 9th IEEE International Conference on Computer Vision. Nice, France:IEEE, 2003. 750-757 [20] 韩贵金, 朱虹.一种基于图结构模型的人体姿态估计算法.计算机工程与应用, 2013, 49(14):30-33 doi: 10.3778/j.issn.1002-8331.1302-0153Han Gui-Jin, Zhu Hong. Human pose estimation algorithm based on pictorial structure model. Computer Engineering and Applications, 2013, 49(14):30-33 doi: 10.3778/j.issn.1002-8331.1302-0153 [21] Jiang H. Human pose estimation using consistent max covering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(9):1911-1918 doi: 10.1109/TPAMI.2011.92 [22] Yang W L, Wang Y, Mori G. Recognizing human actions from still images with latent poses. In:Proceedings of the 2010 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). San Francisco, USA:IEEE, 2010. 2030-2037 [23] 徐峰, 张军平.人脸微表情识别综述.自动化学报, 2017, 43(3):333-348 http://www.aas.net.cn/CN/abstract/abstract19013.shtmlXu Feng, Zhang Jun-Ping. Facial microexpression recognition:a survey. Acta Automatica Sinica, 2017, 43(3):333-348 http://www.aas.net.cn/CN/abstract/abstract19013.shtml [24] 徐渊, 许晓亮, 李才年, 姜梅, 张建国.结合SVM分类器与HOG特征提取的行人检测.计算机工程, 2016, 42(1):56-60, 65 doi: 10.3969/j.issn.1000-3428.2016.01.011Xu Yuan, Xu Xiao-Liang, Li Cai-Nian, Jiang Mei, Zhang Jian-Guo. Pedestrian detection combining with SVM classifier and HOG feature extraction. Computer Engineering, 2016, 42(1):56-60, 65 doi: 10.3969/j.issn.1000-3428.2016.01.011 [25] Bo L F, Sminchisescu C. Twin gaussian processes for structured prediction. International Journal of Computer Vision, 2010, 87(1-2):28-52 doi: 10.1007/s11263-008-0204-y [26] Cristianini N, Shawe-Taylor J, Elisseeff A, Kandola J. On kernel-target alignment. In:Proceedings of the 14th International Conference on Neural Information Processing Systems:Natural and Synthetic. Vancouver, British Columbia, Canada:MIT Press, 2001. 367-373 [27] Gretton A, Bousquet O, Smola A J, Schölkopf B. Measuring statistical dependence with Hilbert-Schmidt norms. Algorithmic Learning Theory. Berlin Heidelberg, Germany:Springer-Verlag, 2005. 期刊类型引用(3)
1. 岳振宇，范大昭，董杨，纪松，李东子. 一种星载平台轻量化快速影像匹配方法. 地球信息科学学报. 2022(05): 925-939 . 
百度学术2. 王若兰，潘万彬，曹伟娟. 图像局部区域匹配驱动的导航式拼图方法. 计算机辅助设计与图形学学报. 2020(03): 452-461 . 百度学术3. 胡敬双，聂洪玉. 灰度序模式的局部特征描述算法. 中国图象图形学报. 2017(06): 824-832 . 百度学术其他类型引用(9)
-
下载:
计量
- 文章访问数: 2900
- HTML全文浏览量: 352
- PDF下载量: 687
- 被引次数: 12
