-
摘要: 根据研究重点的不同, 从群目标跟踪的建模和滤波算法方面展开综述, 主要包括:量测处理、扩展外形建模、内部关系建模、群分裂/合并建模以及滤波算法等.最后, 基于群目标跟踪现有研究进展和未来可能面临的挑战, 对群目标跟踪领域需要重点研究和关注的方向作了展望.Abstract: On the basis of the research focuses of group target tracking, an overview is presented, which includes modeling and filtering. Special attention is paid to the following areas: measurement processing, modeling of extension shape, modeling inter relationships, modeling of group splitting/combination, and filters for group target tracking based on the aforementioned models. Finally, based on the research progress and the challenges that maybe faced in the future, an outlook is made over the key issues that need to be focused on in the field of group target tracking.
-
Key words:
- Group target /
- random finite set /
- random matrix /
- evolving networks model /
- group refined tracking /
- group splitting/combination
-
行人再识别是指在无重叠视域监控系统中, 检测和匹配在不同视域摄像头下两个行人是否为一个人的任务.广泛应用于公安实时监控区域、捉捕嫌疑人等任务当中, 对维护国家的治安、提高公安办案效率有着重要的意义.但是由于不同视域下的摄像头之间的行人存在分辨率、视角、光照条件、遮挡、背景干扰、行人姿态以及摄像头成像质量的差异, 导致相同的人在不同视域摄像头下存在很大的外观差异, 会给行人再识别带来很大挑战.深度学习在计算机视觉上取得的巨大成功, 使得在行人再识别领域的研究也日益增加, 卷积神经网络(Convolutional neural network, CNN)作为特征提取器用来自主学习特征, 文献[1]提出了利用Siamese网络进行有监督的学习, 来匹配和区分行人对.文献[2]将长短时记忆网络(Long short-term memory, LSTM)引入Siamese网络, 对行人图像进行分割, 用LSTM依次捕捉各区域之间的空间关系, 增强网络的判别能力. Liu等[3]提出一个端到端的基于比较性注意力网络, 使用LSTM来循环生成局部注意力的特征, 提取到更多局部的辨别性信息, 有效提高行人再识别算法性能.由于实际环境条件的需求, 基于视频的行人再识别工作越来越得到关注. Wang等[4]提出的一个区分视频片段的方法(Discriminative video fragments selection and ranking, DVR)框架用于行人再识别, 使用判别性的时空特征选择来自动发现和利用最可靠的视频片段.文献[5]使用自适应的Fisher判别分析来解决视频行人再识别问题.文献[6]利用视频三元组在训练视频中学习视频内距离度量和视频间距离度量, 从而提高识别的准确度.文献[7]建立了时空人体动作模型, 可以为视频中的行人构建很好的时空表示. McLaughlin等[8]提出了卷积神经网络捕捉图像特征, 循环神经网络(Recurrent neural networks, RNN)捕捉帧与帧的时空关系特征来配合Siamese网络, 取得了很好的实验结果.文献[9]使用LSTM网络以循环的方式来聚合每帧的行人特征.文献[10]采用端到端的双流网络, 提取了行人图片特征和运动上下文信息来提高识别率.文献[11]在文献[8]的基础上分别添加了CNN和RNN上的注意力机制, 使其在时空上有重点的关注某些重要特征, 提升了识别效率.
2014年, Goodfellow等[12]提出了生成对抗网络(Generative adversarial networks, GAN).生成对抗网络来源于二人零和博弈思想的启发, 结构分为两个部分, 生成网络(Generator network, G)和判别网络(Discriminator network, D), 生成网络主要是通过捕捉训练集上的数据, 来产生新的样本, 而判别网络则是判断样本是否是生成网络生成的还是来自原训练集.两种网络相互竞争, 最后判别器将无法区分训练数据分布和生成数据分布, 生成的图片满足原训练集的图片分布, 则完成训练, 实验表明在伪造图片方面, GAN网络具有很强的优势.文献[12]的提出使得越来越多的基于生成对抗网络的研究被提出, 文献[13]将GAN扩展到了CNN的领域, 使得GAN的训练更加稳定和可控.文献[14]在原始GAN的基础上加入了监督的信息, 使得生成的图片向着标签方向生成.文献[15]采用序列化的思想, 结合图像拉普拉斯金字塔实现序列化的生成, 减少GAN每次学习的内容和难度, 图像质量得到提升.为了解决GAN模型中存在训练困难、生成器和判别器的损失函数无法指示训练进程、生成样本缺乏多样性等问题, 文献[16]采用Earth-Mover距离替代了JS (Jensen-Shannon)散度来衡量距离, 在近似最优判别器下优化生成器缩小Wasserstein距离, 拉近生成分布与真实分布.
人体姿态估计是通过给定图像来确定图像中各个人体部位的位置的过程, 用来分析人体动作和行为.传统的方法分为基于人体特征[17-19]和基于模型的方法[20].文献[17]使用形状上下文特征作为人体的外观特征, 利用正则最小二乘法和支持向量机来进行回归.文献[18]同样用形状上下文作为行人特征, 通过距离度量来判断图像间的相似度, 用来分析行人姿态估计.文献[19]用Gist特征来作为人体特征, 通过进行非线性近邻元分析图像相似度来确定人体手和头部的姿态.由于人体非刚体的特性, 基于模型的方法被提出, 文献[20]设想人体不同部位是相互独立存在的, 并对不同部位添加了约束, 构造了人体的树形图结构模型作为姿态估计的模型.基于深度学习的姿态估计方法近几年也被提出, 并取得了不错的成绩.文献[21]用卷积神经网络来提取人体特征并计算关节点的分布, 结合现有的人体模型进行姿态估计.文献[22]利用CNN来估计人体姿态, 融合空间信息和光流信息, 并且用热力图(Heatmap)来取代关节点的坐标, 提高了关节点检测的鲁棒性.
目前在行人再识别课题研究中, 大量行人图像重构和生成的行人再识别方法被提出.高质量的行人图片和大量的标签样本有助于网络性能的提高.文献[23]提出了一种半耦合低秩判别字典学习(Semi-coupled low-rank discriminant dictionary learning, SLD$^2$L)方法, 将超分辨率重建引入行人再识别研究. Zheng等[24]将生成对抗网络应用于行人再识别方向, 通过生成对抗网络生成符合行人特征分布的图像, 并平滑行人标签, 提高了匹配效果. Qian等[25]利用生成对抗网络生成出不同姿势的行人图片, 解决了应用场景下行人姿势的干扰, 提升了算法识别效果.为了解决在一个数据集上训练的网络不能应用在另外一个数据集的问题, 文献[26]设计了SPGAN (Similarity preserving cycle-consistent generative adversarial network), 可以无监督地将标记图像从源域转换为目标域.
我们相信在实验中添加带有标签的样本可以提升网络性能, 提高识别效果.因此本文提出了利用生成对抗网络来生成带有标签的视频样本通过增加样本数来提高网络能力, 同时在输入中增加了行人关节点的一维特征.实验显示, 基于文献[11]的方法, 相较于其他行人再识别相关方法, 在PRID2011和iLIDS-VID基于视频的行人再识别数据集上实验, 行人匹配率得到了显著的提升.
本文的其余章节组织安排如下.第1节介绍本文提出的融合生成对抗网络和姿态估计的视频行人再识别方法；第2节介绍本文算法在视频行人再识别公共数据集上的实验；第3节总结全文以及展望.
1. 融合生成对抗网络和姿态估计的视频行人再识别方法
1.1 通过生成对抗网络预测帧以增加样本
文献[27]采用了一种图像多尺度结构(见图 1)结合生成对抗网络对抗训练方法, 并设计了图像梯度差损失函数来保证生成帧的清晰度.为了避免采样和池化所带来的分辨率上的损失, 采用了拉普拉斯金字塔的结构, 通过4次升采样不断的逼近真实样本.拉普拉斯金字塔中各层生成网络生成的预测图像可以表示为
$$ \begin{align} \hat{Y}_{k} = G_{k}(X) = u_k(\hat{Y}_{k-1})+G^{'}_k(X_k, u_k(\hat{Y}_{k-1})) \end{align} $$ (1) 其中, $ k $下标表示不同的输入图片尺寸, $ k $取值从1到4, 分别代表输入尺寸为$ 4\times 4 $, $ 8\times 8 $, $ 16\times16 $和$ 32 $ $ \times $ $ 32 $, $ u_{k} $表示图像升采样到$ k $尺寸大小的图像, $ G^{'}_k $表示输入图像大小符合$ k $尺寸的生成模型, $ \hat{Y}_{k} $表示生成模型生成的符合$ k $尺寸的预测图像.判别模型$ D $输入一系列的图像帧, 用来训练分辨序列的最后一帧是真实的图像还是$ G $生成的图像.同样为了满足生成模型中不同尺寸的变化, $ D $是具有单个标量输出的多尺度卷积网络. $ G, D $交替训练, 使用随机梯度下降算法(Stochastic gradient descent, SGD)来最小化损失函数.判别模型的损失函数$ L_{\rm adv}^D $表示为
$$ \begin{align} L_{\rm adv}^D(X, Y) = & \sum\limits_{k = 1}^{N_{\rm scales}}\big[L_{\rm bce}(D_{k}(X_{k}, Y_{k}), 1)\, +\\ &\ L_{\rm bce}(D_{k}(X_{k}, G_{k}(X)), 0)\big] \end{align} $$ (2) 其中, $ (X, Y) $是来自数据集的样本. $ X $是$ m $个帧的序列, $ Y $是$ X $下一帧的图像.对每个输入的尺寸$ k, $固定生成模型$ (G) $, 对$ D $进行SGD迭代, 训练判别模型可以将真实输入分类到1类, 伪造输入分类到0类. $ L_{\rm bce} $表示的是二元交叉熵损失函数, 表示为
$$ \begin{align} L_{\rm bce}(Y, \hat{Y}) = \! -\sum\limits_{i}\big[\hat{Y_{i}}\lg(Y_{i})+ (1-\hat{Y_{i}})\lg(1-Y_{i})\big] \end{align} $$ (3) 生成模型的损失函数$ L_{\rm adv}^{G} $表示为
$$ \begin{align} L_{\rm adv}^{G}(X, Y) = \sum\limits_{k = 1}^{N_{\rm scales}}L_{\rm bce}(D_{k}(X_{k}, G_{k}(X_{k})), 1) \end{align} $$ (4) 其中, $ (X, Y) $同式(2).固定判别模型$ (D) $, 使用SGD对$ G_{k} $进行迭代以最小化损失函数, 在生成模型中, 最小化损失函数的目的就是生成的图片尽可能地去迷惑判别模型, 使判别模型不能正确地区分真实图片和生成图片.但是在实验中, 这种损失函数的构建会导致系统不稳定, 判别网络会生成混乱的样本来欺骗判别网络, 因此设计联合损失函数$ L(X, Y) $, 在生成模型损失函的基础上加入$ L_{p} $损失函数以及图像梯度差分损失$ L_{\rm gdl} $去联合优化生成网络, 表示为
$$ \begin{align} L(X, Y) = &\ \lambda_{\rm adv}L_{\rm adv}^{G}(X, Y)+\lambda_{l_{p}}L_{p}(X, Y)\, +\\ &\ \lambda_{\rm gdl}L_{\rm gdl}(X, Y) \end{align} $$ (5) 其中, $ \lambda_{\rm adv} $, $ \lambda_{l_{p}} $和$ \lambda_{\rm gdl} $分别为$ L_{\rm adv}^{G} $, $ L_{p} $和$ L_{\rm gdl} $损失函数的权重, $ L_{p}(X, Y) $通过最小化与真实图片的距离来优化模型, $ p $的值可以是1或者是2, $ L_{p} $表示为
$$ \begin{align} L_{p}(X, Y) = l_{p}(G(X), Y) = \left\|G(X)-Y\right\|_{p}^{p} \end{align} $$ (6) $ L_{\rm gdl} $表示为
$$ \begin{align} &L_{\rm gdl}(X, Y) = L_{\rm gdl}(\hat{Y}, Y) = \\ &\qquad\sum\limits_{i, j}\left[\left|\left| Y_{i, j}-Y_{i-1, j} \right|-|{\hat{Y}_{i, j}-\hat{Y}_{i-1, j} }|\right|^a +\right.\\ &\qquad\left.\left|{\left| {Y_{i, j-1}-Y_{i, j}} \right| - |{\hat{Y}_{i, j-1}-\hat{Y}_{i, j}}|}\right|^a\right] \end{align} $$ (7) 式(7)是计算生成图片和真实图片像素之间梯度差异, 其中, $ a $是大于或者等于1的整数, $ |{\cdot}| $表示绝对值函数.加入梯度差分损失函数作为惩罚项, 目的是使生成图像更加锐利.
实验中通过生成对抗网络递归生成所有行人视频序列下一帧的图像, 并利用循环递归来生成后$ N $帧的图像, 并给每帧图像打上该行人身份标签. 图 2中, 每行的前5帧图像是视频行人再识别数据集上的连续图像, 而每行的后5帧是通过前5帧的输入由生成对抗网络生成出来的连续图像.可以看出, 通过训练, 生成的后5帧图像保持了行人正常的步伐轨迹, 同时有较好的图像效果.通过后面实验论证, $ N $值取5帧时算法性能最优, 本文实验统一设$ N $ $ = $ $ 5. $实验中生成图像进行迭代的过程中误差的不断叠加, 导致迭代5帧之后的图片质量难以保证, 会影响行人正确的识别和分类.
图 2 生成对抗网络生成的视频帧序列(后5帧)Fig. 2 A sequence of video frames generated by GAN (last five frames)1.2 行人关节点的检测
目前视频行人再识别的特征主要是神经网络通过原图和光流图来提取特征, 原图可以提取到行人空间上的深度特征, 光流是图像序列中像素在时间域上的变化以及相邻帧之间的相关性来找到上一帧与当前帧之间存在的对应关系的时间特征.我们认为加入行人关节点的一维有效特征会帮助网络正确分类, 因为通过视频中行人关节点的变化情况, 可以提取到关节点运动的时间信息帮助网络分类, 也会让网络关注关节点周围的信息, 避免了背景所带来的影响.
文献[28]在文献[29]的基础上设计了卷积姿态机(Convolutional pose machines, CPM)算法, 算法利用了全卷积的网络学习图像特征和图像间的空间信息来解决姿态估计任务.在任务中设关节点数目为$ p $, 关节点的坐标用$ Y_{p} $来表示, $ \mathcal{Z} $表示所有关节点坐标的集合, 存在$ \forall Y_{p} \in \mathcal{Z} $, 算法目的是预测出$ p $个点的坐标$ Y_{p} $.
图 3 (a)中使用了深度卷积网络来进行局部预测, 使用了五个常规卷积层和两个$ 1\ast 1 $的卷积层构成的网络结构, 并在网络的最后将局部图回归到一个$ p+1 $大小的输出矢量, 表示了图像区域每个关节点的得分信息.图 3 (b)中展示了后续网络的结构, 网络不仅仅会得到原图的特征也会加入上一阶段输出的特征来表示图像上下文之间的联系.并且通过大尺度的卷积($ 11\ast 11 $)增加检测图像感受野的尺度, 使得关节点的检测既能确保精度, 又考虑了各个关节点之间的距离关系.作为一种多阶段的网络结构, 为了防止网络训练梯度传播过程中发生的梯度消失现象, 网络会在每个阶段内添加监督信息, 有效地加强反向传播能力, 避免发生梯度消失.每个阶段优化任务表示为
$$ \begin{align} f_{t} = \sum\limits_{p = 1}^{P+1}\sum\limits_{z\in\mathcal{Z}}\left\|{b_{t}^{p}(z)-b_{*}^{p}(z)}\right\|_{2}^{2} \end{align} $$ (8) 其中, $ b_{t}^{p} $表示CPM网络结构每个阶段$ t $输出的关节点$ p $的置信图, $ b_{*}^{p} $表示原图中对图像标注中的真实关节点$ p $做高斯分布产生.而损失函数在每个阶段$ t $的优化函数是每个像素间的平方误差.
整个CPM网络优化函数表示如式(9), 即每个阶段$ t $优化函数的叠加.网络使用标准随机梯度下降来优化网络性能.
$$ \begin{align} \mathcal{F} = \sum\limits_{t = 1}^{T}f(t) \end{align} $$ (9) 文献[30]基于CPM算法设计了一种复杂场景下实时多人姿态检测的方法, 解决了运算效率低和多人分类不准的问题, 实验在MPII和MSCOCO Keypoint数据集上都展现了很好的效果.网络结构主要由部位置信图(Part confidence maps, PCM)和部位亲和域(Part affinity fields, PAF)组成, 利用PCM来回归关节点和PAF回归出两个关节点之间的热图, 最后通过图论中偶匹配的方式, 连接两个关节点, 获取多人的姿态.
目前的行人再识别数据集中还没有关节点的标签, 利用文献[30]的算法在MSCOCO Keypoint数据集上的训练, 可以获得多达18个关节点的行人关节点特征.如图 4所示, 上列是行人连续运动的视频帧序列, 下列则是根据上列的图像来可视化的人体关节点的一维图像信息, 可以看出很好的检测并可视化出行人18个关节点信息, 描述出行人动作和姿态的特征.
1.3 本文算法的具体步骤
步骤 1. 输入视频行人再识别的数据集, 利用多尺度生成对抗网络去预测下一帧的图像, 用第1.1节中式(5)作为优化任务, 循环递归生成$ N $帧图像, 生成的图像将打上该序列中行人身份标签.
图 5 融合生成对抗网络和姿态估计算法网络结构Fig. 5 The structure of integration of GAN and pose estimation algorithm步骤 2. 利用第1.2节中图 3的网络模型联合式(8)和式(9)的优化任务对扩充的视频行人再识别数据集进行人体姿态估计, 获得数据集中不同行人关节点的一维信息特征.
步骤 3. 利用步骤1扩充的行人再识别数据集和步骤2生成的数据集中人体关节点的信息, 与YUV颜色三通道和光流两通道组成6通道图像, 使用时-空注意力网络(Attentive spatial-temporal pooling networks, ASTPN)进行分类, 获取实验结果.
2. 实验测试与结果
本节首先介绍了行人再识别研究中算法性能的评测准则和实验中所使用的数据集, 其次介绍了本文算法在不同公共实验数据集上的实验结果, 并与已有的行人再识别算法实验性能进行了比较, 分析了两个创新部分分别对算法性能产生的影响.最后本节分析了由生成对抗网络递归产生的帧数对算法性能的影响, 得出的结论使算法性能达到最优.文中的实验通过Torch和TensorFlow框架实现, 硬件采用搭载i5-4590 (3.30 GHz)和NVIDIA GTX-980TI (4 GB)的个人电脑.
2.1 测试数据和算法性能的评测准则
本文在PRID2011和iLIDS-VID视频行人再识别数据集上进行实验与测试.基于文献[11]的描述, 将视频行人再识别数据集上的行人对随机分为两个大小相同的子集, 一个为训练集.另一个为测试集.由于在视频行人再识别数据集上查询集(Probe)和行人图像集(Gallery)序列是长度不固定的, 所以在训练过程中每个训练过程随机地选择$ k $ $ = $ $ 16 $个连续帧的子序列.在测试过程中, 我们将第1台相机视为查询集, 将第2台相机视为行人图像集.并给同一个行人来自摄像机1的子序列和来自摄像机2的子序列打上正确行人对的标签, 而负行人对标签来自不同人在摄像机1和摄像机2的子序列.正标签对和负标签对经过图像增强会被连续地送入ASTPN网络, 使模型能够区分正确匹配和错误匹配.实验采用累积匹配特征曲线(Cumulative match characteristics, CMC)来评价算法的性能. CMC在实验中表示在不同摄像机下行人匹配的概率.实验中给定一个查询的数据集和行人图像数据集(摄像机1和摄像机2), 累积匹配特征曲线描述的是在行人图像数据集中搜索待查询数据集中的行人, 前$ r $个搜索结果中显示找出待查询人查询到的比率.在日常生活中, 我们人类衡量匹配的准确率往往是第一匹配率(Rank 1), 因此我们在实验中同样将第一匹配率放在衡量算法性能标注的首要位置.每组实验将训练集和测试集过程重复10次, 取10次的平均值作为该组实验的结果.
2.2 在不同数据集上的实验结果
2.2.1 PRID2011数据集
PRID2011数据集中有385个行人视频序列来自摄像机$ A $, 有749个行人视频序列来自摄像机$ B $.但是只有200个人同时出现在两台摄像机中, 每个行人序列的长度在5到675个图像帧不等, 平均行人视频帧数为100帧, 整个数据集总共有24 541张图片, 图片尺寸为$ 128 $像素$ \times $ $ 64 $像素大小.
表 1给出了本文改进方法针对PRID2011数据集相对于其他算法累积特征匹配曲线的对比, 可以看出本文改进方法相较于ASTPN和其他基于视频的行人再识别方法来说, 可以有效地提升CMC中Rank 1的识别率, 对Rank 5的识别率也有提升.对比ASTPN算法, 加入更多样本和特征的信息会让Rank 1提升了3.2 %.
表 1 不同算法在PRID2011数据集上的识别率(%)Table 1 Matching rates of different methods on the PRID2011 dataset (%)从表 2可以看出, 在PRID2011数据集上通过生成对抗网络生成更多带标签的样本和添加了行人关节点的特征信息, 都对行人再识别的Rank 1有所提高, 分别提升2.2 %和1.6 %.
表 2 不同算法在PRID2011数据集上对识别率的影响(%)Table 2 The influence of different methods on matching rates based on PRID2011 dataset (%)方法 Rank 1 Rank 5 Rank 10 Rank 20 ASTPN 77.0 95.0 99.0 99.0 ASTPN+GAN 79.2 95.3 99.2 99.2 ASTPN+KeyPoint 78.6 95.1 99.1 99.1 本文方法 80.2 96.0 99.1 99.2 2.2.2 iLIDS-VID数据集
iLIDS-VID数据集是从机场的两个无重叠视域的摄像机中捕捉到的行人轨迹图像.数据集中包含了300个不同行人的600个图像视频序列, 相同行人的两个视频序列分别采集于摄像机1和摄像机2.在数据集中每个行人视频序列的长度在23到192个图像帧不等, 每个视频序列的平均帧数为73帧, 数据集中总共有42 495张图片.数据集中包含了人与人之间的服装相似性, 不同摄像机光线和视角的变化, 复杂的背景以及存在的遮挡等问题, 给研究人员带来了很大的挑战.
由表 3可知, 本文改进方法对iLIDS-VID数据集同样性能上有所提高, 将Rank 1提高到66.3%.相较于ASTPN有效提高4.3%, Rank 5也提升了2.4%.
表 3 不同算法在iLIDS-VID数据集上的识别率(%)Table 3 Matching rates of different methods on the iLIDS-VID dataset (%)从表 4可以看出, 在iLIDS-VID数据集上, 改进的两种方法都对算法性能有所提升, 分别将Rank 1提升2.4 %和2.5 %, Rank 5提升1.5 %和1.5 %.
表 4 不同算法在iLIDS-VID数据集上对识别率的影响(%)Table 4 The influence of different methods on matching rates based on iLIDS-VID dataset (%)方法 Rank 1 Rank 5 Rank 10 Rank 20 ASTPN 62.0 86.0 94.0 98.0 ASTPN+GAN 64.4 87.5 95.1 98.0 ASTPN+KeyPoint 64.5 87.5 96.1 98.1 本文方法 66.3 88.4 96.2 98.1 2.3 生成对抗网络递归生成$ {\pmb N} $张图片对方法性能的比较
实验发现, 通过生成对抗网络对视频行人再识别的生成中, 每个行人序列所生成的样本数对行人再识别的累积特征匹配曲线有影响。PRID2011数据集中, 有200对有效视频行人再识别序列, $ N $是每个序列生成对抗网络循环递归产生图片的数量, 数据集会产生$ {N\times 400} $张图片.同理, iLIDS-VID数据集中有300对有效视频行人再识别序列, 故数据集会产生$ {N\times 600} $张图片.表 5和表 6分别给出了本文改进方法和$ N $的大小对PRID2011和iLIDS-VID数据集算法性能影响, 并都加入了行人关节点的一维信息.
表 5 每个行人轨迹递归生成的图片张数$N$对PRID2011数据集上识别率的影响(%)Table 5 The influence of the number $N$ of pictures generated recursively by each pedestrian trace on matching rate based on PRID2011 dataset (%)$ N$ Rank 1 Rank 5 Rank 10 Rank 20 1 77.4 95.2 99.0 99.0 3 78.6 95.6 99.1 99.1 5 80.2 96.0 99.2 99.3 7 80.1 96.0 99.2 99.2 9 77.6 95.5 98.7 99.1 11 76.7 95.4 98.2 99.0 表 6 每个行人轨迹递归生成的图片张数$N$对iLIDS-VID数据集上识别率的影响(%)Table 6 The influence of the number $N$ of pictures generated recursively by each pedestrian trace on matching rate based on iLIDS-VID dataset (%)$N $ Rank 1 Rank 5 Rank 10 Rank 20 1 61.9 86.7 94.2 97.8 3 63.1 87.5 95.6 98.0 5 66.3 88.4 96.2 98.1 7 66.0 88.4 96.0 98.1 9 64.8 87.9 94.3 97.9 11 64.6 86.6 94.1 97.6 从表 5和表 6可知, 当$ N = 1 $到$ N = 5 $之间, 算法性能会逐渐上升, 在$ N = 5 $和$ N = 7 $之间算法性能达到最优, 而在$ N>7 $之后, 算法性能逐渐下降, 考虑到算法效率结合最优结果, 本实验会将$ N $值设为5.通过实验验证, 每个行人视频序列生成的$ N $并不是越多越好, 由于生成对抗网络递归中误差的不断叠加, 在后续产生的图像中行人特征会变差, 干扰了网络特征提取和正确分类.
3. 结束语
现有的基于视频的行人再识别方法, 主要从设计提取鲁棒的时间空间特征深度网络和设计泛化性能好的分类器这两个方向入手.但是现有视频行人再识别数据集一般规模较小, 没有充分的监督样本来进行网络的训练和优化, 容易使深度网络发生过拟合的现象, 泛化性不高.对此, 本文提出融合生成对抗网络和姿态估计的视频行人再识别方法.该方法首先利用多尺度生成对抗网络去生成行人带标签的视频序列, 生成的视频序列无论行人外貌和运动轨迹都接近于真实视频序列, 有效扩大了视频行人再识别数据集的数据量; 其次加入了视频行人关节点的信息, 增加了行人的姿态特征; 最后利用了时–空注意力网络对处理过的数据集进行分类, 判断样本之间的相似性.两个视频行人再识别数据集上的实验相较于其他改进算法也取得了较好的实验结果.针对生成的图像序列数目有限的问题, 下一步需要解决在多次迭代中图像误差逐渐增大的问题, 产生更多连续高质量图像来扩增数据集.
-
图 3 多椭圆建模任意外形群目标示意图
Fig. 3 Modeling arbitrary shape group targets with multiple ellipse center
表 2 常见群扩展外形的主要建模方法及特点
Table 2 Main modeling methods and features of common group extensions
扩展外形 建模方法 特点 椭圆外形 RM方法[14, 44, 55-57] 波方程简单, 预设先验参数少、鲁棒性强, 适用于线性系统 RHM方法[43-44] 精度高, 比RM方法计算复杂 矩形外形 矩形参数法[11] 波方程简单, 预设先验参数少、鲁棒性强, 适用于线性系统 区间箱方法[46] 密集杂波条件下效果较好; 只描述了量测外形而没有估计目标外形 星凸外形 星凸RHM方法[20, 58] 比简单外形建模方法具有更高的跟踪精度和更强的鲁棒性, 可描述更复杂的群外形; 估计半径可能为负 GP方法[49, 59] 半径函数具有描述后验分布的能力; 不需要进行频域转换; 能保持不可观测部分的不确定性; 能够估计目标方向; 估计半径可能为负 任意外形 水平集RHM方法[40] 能够在信息不确定条件下对时变外形进行高精度跟踪; 初始化困难; 需要正则化以满足鲁棒性要求; 需要处理强噪声造成的估计偏差 多椭圆方法[12-13, 16, 52] 利用多个简单外形描述复杂外形, 具有较高的灵活性, 算法复杂度随着子目标数目的增加而增加 
下载: 导出CSV
表 3 常见群内部关系模型及特点
Table 3 Common group internal relationship models and their characteristics
下载: 导出CSV
-
[1] Beard M, Reuter S, Granström K, Vo B T, Vo B N, Scheel A. Multiple extended target tracking with labeled random finite sets. IEEE Transactions on Signal Processing, 2016, 64(7): 1638-1653 doi: 10.1109/TSP.2015.2505683 [2] 耿文东, 王元钦, 董正宏.群目标跟踪.北京:国防工业出版社, 2014.Geng Wen-Dong, Wang Yuan-Qin, Dong Zheng-Hong. Group-Targets Tracking. Beijing: National Defense Industry Press, 2014. [3] Mihaylova L, Carmi A Y, Septier F, Gning A, Pang S K, Godsill S. Overview of Bayesian sequential Monte Carlo methods for group and extended object tracking. Digital Signal Processing, 2014, 25: 1-16 doi: 10.1016/j.dsp.2013.11.006 [4] Waxman M J. A bibliography of cluster (Group) tracking. In: Proceedings of the SPIE 5428, Signal and Data Processing of Small Targets. Florida, USA: SPIE, 2004. 551- 560 [5] Drummond O E, Blackman S S, Pretrisor G C. Tracking clusters and extended objects with multiple sensors. In: Proceedings of the SPIE 1305, Signal and Data Processing of Small Targets. Florida, USA: SPIE, 1990. 362-375 [6] 李振兴.群目标跟踪关键技术研究[博士学位论文], 空军工程大学, 中国, 2014.Li Zhen-Xing. Research on the Key Technologies of Group Target Tracking [Ph. D. dissertation], Air Force Engineering University, China, 2014. [7] 王海鹏, 董云龙, 熊伟, 潘新龙, 贾舒宜.多传感器编队目标跟踪技术.北京:电子工业出版社, 2016.Wang Hai-Peng, Dong Yun-Long, Xiong Wei, Pan Xin-Long, Jia Shu-Yi. Multi-Sensor Formation Targets Tracking Method. Beijing: Electronic Industry Press, 2016. [8] Taenzer E. Tracking multiple targets simultaneously with a phased array radar. IEEE Transactions on Aerospace and Electronic Systems, 1980, AES-16(5): 604-614 doi: 10.1109/TAES.1980.308926 [9] Baum M, Faion F, Hanebeck U D. Modeling the target extent with multiplicative noise. In: Proceedings of the 15th International Conference on Information Fusion. Singapore, Singapore: IEEE, 2012. 2406-2412 [10] Koch J W. Bayesian approach to extended object and cluster tracking using random matrices. IEEE Transactions on Aerospace and Electronic Systems, 2008, 44(3): 1042- 1059 doi: 10.1109/TAES.2008.4655362 [11] Granström K, Reuter S, Meissner D, Scheel A. A multiple model PHD approach to tracking of cars under an assumed rectangular shape. In: Proceedings of the 17th International Conference on Information Fusion (Fusion). Salamanca, Spain: IEEE, 2014. 1-8 [12] Lan J, Li X R. Tracking of extended object or target group using random matrix, Part Ⅱ: irregular object. In: Proceedings of the 15th International Conference on Information Fusion. Singapore, Singapore: IEEE, 2012. 2185- 2192 [13] Lan J, Li X R. Tracking of maneuvering non-ellipsoidal extended object or target group using random matrix. IEEE Transactions on Signal Processing, 2014, 62(9): 2450- 2463 doi: 10.1109/TSP.2014.2309561 [14] Lan J, Li X R. Tracking of extended object or target group using random matrix, Part Ⅰ: new model and approach. In: Proceedings of the 15th International Conference on Information Fusion. Singapore, Singapore: IEEE, 2012. 2177-2184 [15] Granstrom K, Baum M, Reuter S. Extended object tracking: introduction, overview and applications. Journal of Advances in Information Fusion, 2017, 12(2): 139-174 http://d.old.wanfangdata.com.cn/NSTLQK/NSTL_QKJJ026837439/ [16] Hu Q, Ji H B, Zhang Y Q. Tracking of maneuvering non-ellipsoidal extended target with varying number of sub-objects. Mechanical Systems and Signal Processing, 2018, 99: 262-284 doi: 10.1016/j.ymssp.2017.06.013 [17] Granström K, Lundquist C, Orguner U. A Gaussian mixture PHD filter for extended target tracking. In: Proceedings of the 13th International Conference on Information Fusion. Edinburgh, UK: IEEE, 2010. 1-8 [18] Granström K, Lundquist C, Orguner U. Extended target tracking using a Gaussian-mixture PHD filter. IEEE Transactions on Aerospace and Electronic Systems, 2012, 48(4): 3268-3286 doi: 10.1109/TAES.2012.6324703 [19] 张永权.随机有限集扩展目标跟踪算法研究[博士学位论文], 西安电子科技大学, 中国, 2014.Zhang Yong-Quan. Research on Algorithms of Extended Target Tracking Based on Random Finite Set [Ph. D. dissertation], Xidian University, China, 2014. [20] Li P, Ge H W, Yang J L. Adaptive measurement partitioning algorithm for a Gaussian inverse Wishart PHD filter that tracks closely spaced extended targets. Radioengineering, 2017, 26(2): 573-580 doi: 10.13164/re.2017.0573 [21] 王聪, 王海鹏, 何友.基于坐标映射距离差分的快速群分割算法.系统工程与电子技术, 2016, 38(8): 1716-1722 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=xtgcydzjs201608002Wang Cong, Wang Hai-Peng, He You. Fast algorithm of group segmentation based on coordinates transformations and distance differentiations. Systems Engineering and Electronics, 2016, 38(8): 1716-1722 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=xtgcydzjs201608002 [22] Granström K, Orguner U. A PHD filter for tracking multiple extended targets using random matrices. IEEE Transactions on Signal Processing, 2012, 60(11): 5657-5671 doi: 10.1109/TSP.2012.2212888 [23] Yang J L, Li P, Yang L, Ge H W. An improved ET-GM-PHD filter for multiple closely-spaced extended target tracking. International Journal of Control, Automation and Systems, 2017, 15(1): 468-472 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=59912f6d99eec7b18579787a001e6356 [24] Zhang Y Q, Ji H B. A novel fast partitioning algorithm for extended target tracking using a Gaussian mixture PHD filter. Signal Processing, 2013, 93(11): 2975-2985 doi: 10.1016/j.sigpro.2013.04.006 [25] Zhang Y Q, Ji H B. Clustering Gaussian mixture reduction algorithm based on fuzzy adaptive resonance theory for extended target tracking. IET Radar, Sonar and Navigation, 2014, 8(5): 536-546 doi: 10.1049/iet-rsn.2013.0254 [26] Zhang Y Q, Ji H B. Gaussian mixture reduction based on fuzzy ART for extended target tracking. Signal Processing, 2014, 97: 232-241 doi: 10.1016/j.sigpro.2013.11.004 [27] Zhang Y Q, Ji H B. A robust and fast partitioning algorithm for extended target tracking using a Gaussian inverse Wishart PHD filter. Knowledge-Based Systems, 2016, 95: 125-141 doi: 10.1016/j.knosys.2015.12.008 [28] 冯新喜, 蒲磊, 孔云波, 王雪.基于随机有限集理论的多扩展目标跟踪技术综述.空军工程大学学报(自然科学版), 2016, 17(3): 93-99 doi: 10.3969/j.issn.1009-3516.2016.03.018Feng Xin-Xi, Pu Lei, Kong Yun-Bo, Wang Xue. A survey of multiple extended targets tracking techniques based on FISST. Journal of Air Force Engineering University (Natural Science Edition), 2016, 17(3): 93-99 doi: 10.3969/j.issn.1009-3516.2016.03.018 [29] Shen X L, Song Z Y, Fan H Q, Fu Q. Generalised distance partitioning for multiple-detection tracking filter based on random finite set. IET Radar, Sonar and Navigation, 2018, 12(2): 260-267 doi: 10.1049/iet-rsn.2017.0201 [30] Granström K, Lundquist C. On the use of multiple measurement models for extended target tracking. In: Proceedings of the 16th International Conference on Information Fusion. Istanbul, Turkey: IEEE, 2013. 1534-1541 [31] Granström K, Orguner U. On spawning and combination of extended/group targets modeled with random matrices. IEEE Transactions on Signal Processing, 2013, 61(3): 678 -692 doi: 10.1109/TSP.2012.2230171 [32] Petrov N, Mihaylova L, Gning A, Angelova D. A novel sequential Monte Carlo approach for extended object tracking based on border parameterisation. In: Proceedings of the 14th International Conference on Information Fusion, Chicago, Illinois, USA: IEEE, 2011. 306-313 [33] Schuster M, Reuter J, Wanielik G. Probabilistic data association for tracking extended targets under clutter using random matrices. In: Proceedings of the 18th International Conference on Information Fusion. Washington, USA: IEEE, 2015. 961-968 [34] Zhu H Y, Han C Z, Li C. An extended target tracking method with random finite set observations. In: Proceedings of the 14th International Conference on Information Fusion. Chicago, USA: IEEE, 2011. 1-6 [35] Baum M, Noack B, Hanebeck U D. Extended object and group tracking with elliptic random hypersurface models. In: Proceedings of the 13th International Conference on Information Fusion. Edinburgh, UK: IEEE, 2010. 1-8 [36] Granström K, Lundquist C, Orguner U. Tracking rectangular and elliptical extended targets using laser measurements. In: Proceedings of the 14th International Conference on Information Fusion. Chicago, USA: IEEE, 2011. 1 -8 [37] Petrov N, Gning A, Mihaylova L, Angelova D. Box particle filtering for extended object tracking. In: Proceedings of the 15th International Conference on Information Fusion. Singapore, Singapore: IEEE, 2012. 82-89 [38] 严超.箱粒子滤波在扩展目标跟踪中的应用[硕士学位论文], 西安电子科技大学, 中国, 2014.Yan Chao. Box Particle Filter's Application to Extended Target Tracking [Master thesis], Xidian University, China, 2014. [39] Baum M, Feldmann M, Franken D, Hanebeck U D, Koch W. Extended object and group tracking: a comparison of random matrices and random hypersurface models. In: Proceedings of the 5th IEEE ISIF Workshop on Sensor Data Fusion: Trends, Solutions, Applications (SDF). Leipzig, Germany: IEEE, 2010. 904-906 [40] Baum M, Hanebeck U D. Shape tracking of extended objects and group targets with star-convex RHMs. In: Proceedings of the 14th International Conference on Information Fusion. Chicago, USA: IEEE, 2011. 1-8 [41] Baum M, Hanebeck U D. Extended object tracking with random hypersurface models. IEEE Transactions on Aerospace and Electronic Systems, 2014, 50(1): 149-159 doi: 10.1109/TAES.2013.120107 [42] Hirscher T, Scheel A, Reuter S, Dietmayer K. Multiple extended object tracking using Gaussian processes. In: Proceedings of the 19th International Conference on Information Fusion (Fusion). Heidelberg, Germany: IEEE, 2016. 868-875 [43] Zea A, Faion F, Baum M, Hanebeck U D. Level-set random hypersurface models for tracking nonconvex extended objects. IEEE Transactions on Aerospace and Electronic Systems, 2016, 52(6): 2990-3007 doi: 10.1109/TAES.2016.130704 [44] Feldmann M, Franken D. Advances on tracking of extended objects and group targets using random matrices. In: Proceedings of the 12th International Conference on Information Fusion. Seattle, USA: IEEE, 2009. 1029-1036 [45] Feldmann M, Franken D, Koch W. Tracking of extended objects and group targets using random matrices. IEEE Transactions on Signal Processing, 2011, 59(4): 1409- 1420 doi: 10.1109/TSP.2010.2101064 [46] Baum M, Hanebeck U D. Random hypersurface models for extended object tracking. In: Proceedings of the 2009 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT). Ajman, United Arab Emirates: IEEE, 2009. 178-183 [47] 李翠芸, 林锦鹏, 姬红兵.一种基于椭圆RHM的扩展目标Gamma高斯混合CPHD滤波器.控制与决策, 2015, 30(9): 1551-1558 http://d.old.wanfangdata.com.cn/Periodical/kzyjc201509003Li Cui-Yun, Lin Jin-Peng, Ji Hong-Bing. A Gamma Gaussian-mixture CPHD filter based on ellipse random hypersurface models for extended targets. Control and Decision, 2015, 30(9): 1551-1558 http://d.old.wanfangdata.com.cn/Periodical/kzyjc201509003 [48] Wahlström N, Özkan E. Extended target tracking using Gaussian processes. IEEE Transactions on Signal Processing, 2015, 63(16): 4165-4178 doi: 10.1109/TSP.2015.2424194 [49] Aftab W, De Freitas A, Arvaneh M, Mihaylova L. A Gaussian process approach for extended object tracking with random shapes and for dealing with intractable likelihoods. In: Proceedings of the 22nd International Conference on Digital Signal Processing (DSP). London, UK: IEEE, 2017. 1-6 [50] Zea A, Faion F, Baum M, Hanebeck U D. Level-set random hypersurface models for tracking non-convex extended objects. In: Proceedings of the 16th International Conference on Information Fusion. Istanbul, Turkey: IEEE, 2013. 1760 -1767 [51] Granström K, Willett P, Bar-Shalom Y. An extended target tracking model with multiple random matrices and unified kinematics. In: Proceedings of the 18th International Conference on Information Fusion. Washington, USA: IEEE, 2015. 1007-1014 [52] Lan J, Li X R. Joint tracking and classification of non-ellipsoidal extended object using random matrix. In: Proceedings of the 17th International Conference on Information Fusion. Salamanca, Spain: IEEE, 2014. 1-8 [53] Cao X M, Lan J, Li X R. Extension-deformation approach to extended object tracking. In: Proceedings of the 19th International Conference on Information Fusion. Heidelberg, Germany: IEEE, 2016. 1185-1192 [54] Cao X M, Lan J, Li X R. Extended object tracking using control-points-based extension deformation. In: Proceedings of the 20th International Conference on Information Fusion. Xi'an, China: IEEE, 2017. 1-8 [55] Cao W, Lan J, Li X R. Extended object tracking and classification using radar and ESM sensor data. IEEE Signal Processing Letters, 2018, 25(1): 90-94 doi: 10.1016/j.sigpro.2017.10.026 [56] Hu Q, Ji H B, Zhang Y Q. A standard PHD filter for joint tracking and classification of maneuvering extended targets using random matrix. Signal Processing, 2018, 144: 352-363 doi: 10.1016/j.sigpro.2017.10.026 [57] Vivone G, Granström K, Braca P, Willett P. Multiple sensor measurement updates for the extended target tracking random matrix model. IEEE Transactions on Aerospace and Electronic Systems, 2017, 53(5): 2544-2558 doi: 10.1109/TAES.2017.2704166 [58] Yao G, Dani A. Image moment-based random hypersurface model for extended object tracking. In: Proceedings of the 20th International Conference on Information Fusion. Xi'an, China: IEEE, 2017. 1-7 [59] Michaelis M, Berthold P, Meissner D, Wuensche H J. Heterogeneous multi-sensor fusion for extended objects in automotive scenarios using Gaussian processes and a GMPHD-filter. In: Proceedings of the 2017 Sensor Data Fusion: Trends, Solutions, Applications (SDF). Bonn, Germany: IEEE, 2017. 1-6 [60] Pang S K, Li J, Godsill S J. Models and algorithms for detection and tracking of coordinated groups. In: Proceedings of the 2008 IEEE Aerospace Conference. Big Sky, USA: IEEE, 2008. 1-17 [61] Khan Z, Balch T, Dellaert F. MCMC-based particle filtering for tracking a variable number of interacting targets. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(11): 1805-1819 doi: 10.1109/TPAMI.2005.223 [62] Tchango A F, Thomas V, Buffet O, Dutech A, Flacher F. Tracking multiple interacting targets using a joint probabilistic data association filter. In: Proceedings of the 17th International Conference on Information Fusion. Salamanca, Spain: IEEE, 2014. 1-8 [63] Pang S K, Li J, Godsill S J. Detection and tracking of coordinated groups. IEEE Transactions on Aerospace and Electronic Systems, 2011, 47(1): 472-502 doi: 10.1109/TAES.2011.5705687 [64] Gning A, Mihaylova L, Maskell S, Pang S K, Godsill S. Group object structure and state estimation with evolving networks and Monte Carlo methods. IEEE Transactions on Signal Processing, 2011, 59(4): 1383-1396 doi: 10.1109/TSP.2010.2103062 [65] Georgieva P, Mihaylova L, Jain L C. Advances in Intelligent Signal Processing and Data Mining. Berlin, Heidelberg: Springer, 2013. 7-53 [66] Chen X J, Qin Z, An L, Bhanu B. Multiperson tracking by online learned grouping model with nonlinear motion context. IEEE Transactions on Circuits and Systems for Video Technology, 2016, 26(12): 2226-2239 doi: 10.1109/TCSVT.2015.2511480 [67] Qin Z, Shelton C R. Social grouping for multi-target tracking and head pose estimation in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(10): 2082-2095 doi: 10.1109/TPAMI.2015.2505292 [68] Bazzani L, Zanotto M, Cristani M, Murino V. Joint individual-group modeling for tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(4): 746-759 doi: 10.1109/TPAMI.2014.2353641 [69] Ur-Rehman A, Naqvi S M, Mihaylova L, Chambers J A. Multi-target tracking and occlusion handling with learned variational Bayesian clusters and a social force model. IEEE Transactions on Signal Processing, 2016, 64(5): 1320-1335 doi: 10.1109/TSP.2015.2504340 [70] Gao S, Chen X G, Ye Q X, Xing J L, Kuijper A, Ji X Y. A graphical social topology model for multi-object tracking [Online], available: https: //arxiv.org/abs/1702.04040, September 29, 2017 [71] Liu W F, Wen C L. The structure-labeled group target estimation with random finite set observation. In: Proceedings of the 2014 International Conference on Control, Automation and Information Sciences (ICCAIS 2014). Gwangju, South Korea: IEEE, 2014. 1-8 [72] Mahler R P S. Statistical Multisource-Multitarget Information Fusion. Norwood: Artech House, Inc, 2007. [73] Feng P M, Wang W W, Dlay S, Naqvi S M, Chambers J. Social force model-based MCMC-OCSVM particle PHD filter for multiple human tracking. IEEE Transactions on Multimedia, 2017, 19(4): 725-739 doi: 10.1109/TMM.2016.2638206 [74] Mazzon R, Cavallaro A. Multi-camera tracking using a multi-goal social force model. Neurocomputing, 2013, 100: 41-50 doi: 10.1016/j.neucom.2011.09.038 [75] Geng W D, Wang Y Q, Dong Z H. Group-Target Tracking. Singapore: Springer, 2017. 115-142 [76] Lian F, Han C Z, Liu W F, Yan X X, Zhou H Y. Sequential Monte Carlo implementation and state extraction of the group probability hypothesis density filter for partly unresolvable group targets-tracking problem. IET Radar, Sonar and Navigation, 2010, 4(5): 685-702 doi: 10.1049/iet-rsn.2009.0109 [77] 汪云, 胡国平, 甘林海.基于多模型GGIW-CPHD滤波的群目标跟踪算法.华中科技大学学报(自然科学版), 2017, 45(2): 89 -94 http://d.old.wanfangdata.com.cn/Periodical/hzlgdxxb201702017Wang Yun, Hu Guo-Ping, Gan Lin-Hai. Group targets tracking algorithm using a multiple models Gaussian inverse Wishart CPHD filter. Journal of Huazhong University of Science and Technology (Natural Science Edition), 2017, 45(2): 89-94 http://d.old.wanfangdata.com.cn/Periodical/hzlgdxxb201702017 [78] Gan L H, Wang G. Tracking maneuvering group target with extension predicted and best model augmentation method adapted. Mathematical Problems in Engineering, 2017, 2017: Article No. 5752870 [79] Lan J, Li X R. Extended object or group target tracking using random matrix with nonlinear measurements. In: Proceedings of the 19th International Conference on Information Fusion. Heidelberg, Germany: IEEE, 2016. 901-908 [80] Lei B T, Li C Y, Ji H B. Nonlinear maneuvering non-ellipsoidal extended object tracking using random matrix. In: Proceedings of the 20th International Conference on Information Fusion. Xi'an, China: IEEE, 2017. 1-6 [81] Özkan E, Wahlström N, Godsill S J. Rao-Blackwellised particle filter for star-convex extended target tracking models. In: Proceedings of the 19th International Conference on Information Fusion. Heidelberg, Germany: IEEE, 2016. 1193-1199 [82] Simon D. Optimal State Estimation: Kalman, H∞, and Nonlinear Approaches. New York: Wiley-Interscience, 2006. [83] Baum M. Simultaneous Tracking and Shape Estimation of Extended Objects [Ph. D. dissertation], Karlsruhe Institute of Technology, Karlsruhe, Germany, 2013. [84] Steinbring J, Baum M, Zea A, Faion F, Hanebeck U D. A closed-form likelihood for particle filters to track extended objects with star-convex RHMs. In: Proceedings of the 2015 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI). San Diego, USA: IEEE, 2015. 25-30 [85] de Freitas A, Mihaylova L, Gning A, Angelova D, Kadirkamanathan V. Autonomous crowds tracking with box particle filtering and convolution particle filtering. Automatica, 2016, 69: 380-394 doi: 10.1016/j.automatica.2016.03.009 [86] Zhang Y Q, Ji H B, Gao X B, Hu Q. An ellipse extended target CBMeMBer filter using gamma and box-particle implementation. Signal Processing, 2018, 149: 88-102 doi: 10.1016/j.sigpro.2018.03.002 [87] Zhang Y Q, Ji H B, Hu Q. A fast ellipse extended target PHD filter using box-particle implementation. Mechanical Systems and Signal Processing, 2018, 99: 57-72 doi: 10.1016/j.ymssp.2017.05.044 [88] Zhu H Y, Ma T T, Chen S, Jiang W. A random matrix based method for tracking multiple extended targets. In: Proceedings of the 17th International Conference on Information Fusion. Salamanca, Spain: IEEE, 2014. 1-8 [89] Streit R. JPDA intensity filter for tracking multiple extended objects in clutter. In: Proceedings of the 19th International Conference on Information Fusion. Heidelberg, Germany: IEEE, 2016. 1477-1484 [90] Wieneke M, Koch W. A PMHT approach for extended objects and object groups. IEEE Transactions on Aerospace and Electronic Systems, 2012, 48(3): 2349-2370 doi: 10.1109/TAES.2012.6237596 [91] Zhang Y Q, Ji H B, Hu Q. A box-particle implementation of standard PHD filter for extended target tracking. Information Fusion, 2017, 34: 55-69 doi: 10.1016/j.inffus.2016.06.007 [92] Li P, Ge H W, Yang J L, Zhang Y Q. Shape selection partitioning algorithm for Gaussian inverse Wishart probability hypothesis density filter for extended target tracking. IET Signal Processing, 2017, 10(9): 1041-1051 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=6b5a1efda6eddfb1faf399e30e6513ef [93] Lundquist C, Granström K, Orguner U. An extended target CPHD filter and a Gamma Gaussian inverse Wishart implementation. IEEE Journal of Selected Topics in Signal Processing, 2013, 7(3): 472-483 doi: 10.1109/JSTSP.2013.2245632 [94] Li C Y, Wang R, Hu Y H, Wang J B. Cardinalised probability hypothesis density tracking algorithm for extended objects with glint noise. IET Science, Measurement and Technology, 2016, 10(5): 528-536 doi: 10.1049/iet-smt.2016.0004 [95] Jiang T Y, Liu M Q, Fan Z, Zhang S L. On multiple-model extended target multi-Bernoulli filters. Digital Signal Processing, 2016, 59: 76-85 doi: 10.1016/j.dsp.2016.08.002 [96] Zong P, Barbary M. Improved multi-Bernoulli filter for extended stealth targets tracking based on sub-random matrices. IEEE Sensors Journal, 2016, 16(5): 1428-1447 doi: 10.1109/JSEN.2015.2499268 [97] Ma D D, Lian F, Liu J. Sequential Monte Carlo implementation of cardinality balanced multi-target multi-Bernoulli filter for extended target tracking. IET Radar, Sonar and Navigation, 2016, 10(2): 272-277 doi: 10.1049/iet-rsn.2015.0081 [98] Li M, Lin Z P, An W, Zhou Y Y. Box-particle labeled multi-Bernoulli filter for multiple extended target tracking. Radioengineering, 2016, 25(3): 527-535 doi: 10.13164/re.2016.0527 [99] Li W L, Jia Y M, Du J P, Yu F S. Gaussian mixture PHD filter for multiple maneuvering extended targets tracking. In: Proceedings of the 50th IEEE Conference on Decision and Control and European Control Conference. Orlando, USA: IEEE, 2014. 2410-2415 [100] 欧阳成, 陈晓旭, 华云.改进的最适高斯近似概率假设密度滤波.雷达学报, 2013, 2(2): 239-246 http://d.old.wanfangdata.com.cn/Periodical/ldxb201302015Ouyang Cheng, Chen Xiao-Xu, Hua Yun. Improved best-fitting Gaussian approximation PHD filter. Journal of Radars, 2013, 2(2): 239-246 http://d.old.wanfangdata.com.cn/Periodical/ldxb201302015 [101] 韩玉兰, 朱洪艳, 韩崇昭, 王静.多扩展目标的高斯混合概率假设密度滤波器.西安交通大学学报, 2014, 48(4): 95-101 http://d.old.wanfangdata.com.cn/Periodical/xajtdxxb201404017Han Yu-Lan, Zhu Hong-Yan, Han Chong-Zhao, Wang Jing. Gaussian-mixture probability hypothesis density filter for multiple extended targets. Journal of Xi'an Jiaotong University, 2014, 48(4): 95-101 http://d.old.wanfangdata.com.cn/Periodical/xajtdxxb201404017 [102] Wang Y, Hu G P, Zhou H. Group targets tracking using multiple models GGIW-CPHD based on best-fitting Gaussian approximation and strong tracking filter. Journal of Sensors, 2016, 2016: Article No. 7294907 [103] Wang Y, Hu G P, Li Z X. Tracking of group targets using multiple models GGIW-PHD algorithm based on best-fitting Gaussian approximation and strong tracking filter. Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 2018, 232(2): 331-343 doi: 10.1177/0954410016684359 [104] 宋骊平, 严超, 姬红兵, 梁萌.基于箱粒子的多扩展目标PHD滤波.控制与决策, 2015, 30(10): 1759-1765 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=kzyjc201510004Song Li-Ping, Yan Chao, Ji Hong-Bing, Liang Meng. PHD filter for tracking multiple extended targets using box particle. Control and Decision, 2015, 30(10): 1759-1765 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=kzyjc201510004 [105] 汪云.群目标跟踪关键技术研究[博士学位论文], 空军工程大学, 中国, 2017.Wang Yun. Research on the Group Target Tracking Algorithm [Ph. D. dissertation], Air Force Engineering University, China, 2017. [106] Ristic B, Sherrah J. Bernoulli filter for joint detection and tracking of an extended object in clutter. IET Radar, Sonar and Navigation, 2013, 7(1): 26-35 doi: 10.1049/iet-rsn.2012.0069 [107] 李鹏涛.基于箱粒子滤波的MeMBer扩展目标跟踪算法研究[硕士学位论文], 西安电子科技大学, 中国, 2016.Li Peng-Tao. Research on MeMBer Extended Target Tracking Based on Box Particle Filter [Master thesis], Xidian University, China, 2016. [108] Lau R A, Williams J L. Tracking a coordinated group using expectation maximisation. In: Proceedings of the 2013 IEEE Eighth International Conference on Intelligent Sensors, Sensor Networks and Information Processing. Melbourne, Australia: IEEE, 2013. 282-287 [109] 朱志宇.粒子滤波算法及其应用.北京:科学出版社, 2010.Zhu Zhi-Yu. Particle Filter and Its Application. Beijing: Science Press, 2010. [110] 杨小军, 潘泉, 王睿, 张洪才.粒子滤波进展与展望.控制理论与应用, 2006, 23(2): 261-267 doi: 10.3969/j.issn.1000-8152.2006.02.019Yang Xiao-Jun, Pan Quan, Wang Rui, Zhang Hong-Cai. Development and prospect of particle filtering. Control Theory and Applications, 2006, 23(2): 261-267 doi: 10.3969/j.issn.1000-8152.2006.02.019 [111] Carmi A Y, Mihaylova L, Septier F, Pang S K, Gurfil P, Godsill S J. Inferring leadership from group dynamics using Markov chain Monte Carlo methods. Modeling, Simulation and Visual Analysis of Crowds. New York: Springer, 2013. 325-346 [112] Carmi A Y, Mihaylova L, Septier F. Subgradient-based Markov chain Monte Carlo particle methods for discrete-time nonlinear filtering. Signal Processing, 2016, 120: 532 -536 doi: 10.1016/j.sigpro.2015.10.015 [113] 李振兴, 刘进忙, 李松, 白东颖, 倪鹏.基于箱式粒子滤波的群目标跟踪算法.自动化学报, 2015, 41(4): 785-798 doi: 10.16383/j.aas.2015.c140222Li Zhen-Xing, Liu Jin-Mang, Li Song, Bai Dong-Ying, Ni Peng. Group targets tracking algorithm based on box particle filter. Acta Automatica Sinica, 2015, 41(4): 785-798 doi: 10.16383/j.aas.2015.c140222 [114] 李振兴, 刘进忙, 李超, 白东颖, 郭相科.联合模糊聚类和拟蒙特卡罗重采样的群目标跟踪算法.控制理论与应用, 2014, 31(11): 1597-1603 doi: 10.7641/CTA.2014.40362Li Zhen-Xing, Liu Jin-Mang, Li Chao, Bai Dong-Ying, Guo Xiang-Ke. Group targets tracking algorithm by combination of fuzzy clustering and Quasi-Monte Carlo resampling method. Control Theory and Applications, 2014, 31(11): 1597-1603 doi: 10.7641/CTA.2014.40362 [115] 朱书军, 刘伟峰, 崔海龙.基于广义标签多伯努利滤波的可分辨群目标跟踪算法.自动化学报, 2017, 43(12): 2178-2189 doi: 10.16383/j.aas.2017.c160334Zhu Shu-Jun, Liu Wei-Feng, Cui Hai-Long. Multiple resolvable groups tracking using the GLMB filter. Acta Automatica Sinica, 2017, 43(12): 2178-2189 doi: 10.16383/j.aas.2017.c160334 期刊类型引用(7)
1. 王吉涛，刘明亮，许森洋，王玺哲，朱强. SWISS整流器分数阶PID控制策略. 哈尔滨理工大学学报. 2024(03): 51-59 . 
百度学术2. 胡春雨，于梦阁，郑旭光，王国晖. 基于扭矩请求的预见性线控油门控制研究. 青岛大学学报(自然科学版). 2023(02): 64-70 . 百度学术3. 黄宇，高珊，李其贤，丁鹏，王东风，申朋宇. SCR脱硝系统的分数阶PI~λD~μ参数优化控制. 动力工程学报. 2022(02): 122-128 . 百度学术4. 刘福才，郭根旺. 气动变载荷加载系统的分数阶PID改进型自抗扰控制. 振动与冲击. 2022(15): 116-121+129 . 百度学术5. 朱敏，臧昭宇，胥子豪，肖阳. 气动调节阀最优分数阶PID控制器设计. 振动与冲击. 2022(22): 267-274 . 百度学术6. 邵林. 基于CMAC神经网络自整定PID在铁矿品位滴定检测中的应用. 佳木斯大学学报(自然科学版). 2022(06): 120-123 . 百度学术7. 刘志臻，芮晖. 基于二自由度预测模型的软扭矩控制方式. 电气传动自动化. 2021(06): 1-5 . 百度学术其他类型引用(5)
-
下载:
计量
- 文章访问数: 6702
- HTML全文浏览量: 2006
- PDF下载量: 945
- 被引次数: 12
