2. 桂林电子科技大学 广西精密导航技术与应用重点实验室, 桂林 541004
2. Guangxi Key Laboratory of Precision Navigation Technology and Application, Guilin University of Electronic Technology, Guilin 541004, China
早期目标跟踪通常假设每个目标在给定的时间步长最多可以产生一个量测值[1]. 利用现代更精确的传感器, 目标可能占据传感器的多个分辨单元, 从而在给定时间步长产生多个量测, 这些目标被称为扩展目标[2]. 扩展目标跟踪的应用包括使用汽车雷达的车辆跟踪、使用地面或者海上雷达站跟踪足够近的飞机或者船舶等.
Mahler首先将随机有限集 (random finite set, RFS) 引入到扩展目标跟踪[3], 将扩展目标状态集和量测数据集均建模为随机有限集, 避免了数据关联问题. Koch提出利用对称正定的随机矩阵 (random matrix, RM) 方法将目标外形建模成椭圆[4], 并首次将贝叶斯方法应用于椭圆外形的扩展目标跟踪, 在忽略传感器误差的假设条件下给出贝叶斯估计的封闭形式解. 文献[5]首先将该方法用于概率假设密度(probability hypothesis density, PHD)滤波器, 提出了基于随机矩阵的扩展目标高斯逆威沙特概率假设密度 (Gaussian inverse Wishart-probability hypothesis density, GIW-PHD) 滤波器, 实现对扩展目标的运动学特征和外形状态估计. 文献[6]将扩展目标量测比建模为伽玛分布, 提出能够实时估计目标量测比率与扩展目标状态的伽玛高斯逆威沙特 (gamma Gaussian inverse Wishart, GGIW) 模型, 文献[7]将GGIW模型应用于扩展目标多伯努利 (multi-Bernoulli, MB) 滤波器. 针对上述滤波器未能直接形成目标轨迹的问题, 文献[8]将广义标签多伯努利算法应用于扩展目标跟踪, 并将目标状态建模为GGIW分布, 通过标签追溯目标的运动轨迹. Granström等人提出了泊松多伯努利混合 (Poisson multi-Bernoulli mixture, PMBM) 算法, 并在扩展目标跟踪中取得比
衍生是扩展目标的重要运动方式, 在军事的体现有战斗机发射导弹、飞机在航母上起飞等. 文献[12]首先提出了扩展目标衍生的研究工作, 基于随机矩阵提出了一种衍生目标状态期望值与原始目标状态期望值相同的衍生模型, 衍生目标的扩展也保持了原始目标扩展的预期值. 当原始目标的外形尺寸与衍生出的目标尺寸不一致时, 这样的简单衍生模型并不适用. 针对该问题, 文献[13]提出多假设模型, 考虑了衍生目标的合理备选方案, 并验证所提的模型提高了GGIW-PHD滤波器对衍生扩展目标的跟踪性能. 之后将学者们多假设模型结合CPHD、
本文在文献[11]的基础上, 提出一种改进的GGIW-PMB滤波算法. 在滤波算法的预测阶段利用多假设模型[12]对衍生事件建模, 根据给定的假设值, 产生多个衍生假设分量, 将衍生分量建模为GGIW形式, 最后在PMB滤波框架下利用量测数据对每个假设分量更新, 实现原始目标和衍生目标的有效跟踪, 增强算法的适用性.
1 扩展目标模型假设在
$ {X_k} = \left\{ {\xi _k^i = \left( {\gamma _k^i, x_k^i, X_k^i} \right)} \right\}_{i = 1}^{{N_{x, k}}} $ | (1) |
其中,
目标运动模型为:
$ x_{k + 1}^i = ({F_{k + 1|k}} \otimes {I_d})x_k^i + \varpi _{k + 1}^i $ | (2) |
其中,
假设
$ {\textit{z}}_k^j = \left( {{H_k} \otimes {I_d}} \right)x_k^i + e_k^j $ | (3) |
其中,
$ {Z_k} = {{\textit{z}}_x}({\textit{z}}) \cup {{\textit{z}}_c}({\textit{z}}) $ | (4) |
其中,
假设
$ \begin{split} p\left( {{\xi _k}} \right) =& GAM\left( {{\gamma _k};{\alpha _k}, {\beta _k}} \right)N\left( {{x_k};{m_k}, {P_k} \otimes {X_k}} \right) \cdot IW\left( {{X_k};{v_k}, {V_k}} \right)\\ =& GGIW\left( {{\xi _k}, {\zeta _k}} \right) \end{split} $ | (5) |
其中,
PMB分为泊松点过程(Poisson point process, PPP)部分和多伯努利(MB)部分, 其中PPP部分表示目标存在但未被传感器检测到的目标, 由符号
假设
$ {D^u}, {\left( {{r^i}, {f^i}\left( \cdot \right)} \right)_{i \in {\mathbb I}}} $ | (6) |
则
$ D_{k|k - 1}^u, {\left( {r_{k|k - 1}^i, f_{k|k - 1}^i\left( \cdot \right)} \right)_{i \in {\mathbb I}}} $ | (7) |
其中,
$ D_{k|k - 1}^u\left( X \right) = {D^b}\left( X \right) + \left\langle {{D^u}, {p_s}{f_{k|k - 1}}} \right\rangle $ | (8) |
$ r_{k|k - 1}^i = \left\langle {{f^i}, {p_s}} \right\rangle {r^i} $ | (9) |
$ f_{k|k - 1}^i = \frac{{\left\langle {{f^i}, {p_s}{f_{k|k - 1}}} \right\rangle }}{{\left\langle {{f^i}, {p_s}} \right\rangle }} $ | (10) |
其中,
假设给出预测步的PMB密度参数如式(7)和量测集合
$ \hat D_{k|k}^u, {\left\{ {\left( {W_{k|k}^j, {{\left\{ {\left( {\hat r_{k|k}^{j, i}, \hat f_{k|k}^{j, i}} \right)} \right\}}_{i \in {\mathbb I}}}} \right)} \right\}_{j \in \mathbb J}} $ | (11) |
其中, PPP部分的更新为:
$ {D^u}\left( {X} \right) = \left( {1 - {p_D}} \right)D_{k|k - 1}^u\left( {X} \right) $ | (12) |
对于之前存在的目标, 每个单目标假设可能包含误检的目标, 或者使用量测更新的目标. 对于误检的目标, 伯努利密度参数为:
$ {r^{i, j}} = \frac{{r_{k|k - 1}^i\left\langle {f_{k|k - 1}^i, \left( {1 - {p_D}} \right)} \right\rangle }}{{1 - r_{k|k - 1}^i + r_{k|k - 1}^i\left\langle {f_{k|k - 1}^i, {p_D}} \right\rangle }} $ | (13) |
$ {f^{j, i}}\left( {X} \right) = \frac{{\left( {1 - {p_D}} \right)\left( {X} \right)f_{k|k - 1}^i\left( {X} \right)}}{{\left\langle {f_{k|k - 1}^i, \left( {1 - {p_D}} \right)} \right\rangle }} $ | (14) |
其中, 单目标假设的更新轨迹对应的伯努利参数为:
$ {r^{j, i}} = 1 $ | (15) |
$ {f^{j, i}}\left( {X} \right) = \frac{{l\left( {X} \right)f_{k|k - 1}^i\left( {X} \right)}}{{\left\langle {f_{k|k - 1}^i, l\left( {X} \right)} \right\rangle }} $ | (16) |
其中,
为了减少多伯努利的数量, 需要从PMBM滤波器中获得保留尽可能多信息的PMB滤波器, 实现该滤波器需要最小化PMBM与PMB之间的KL散度[11], 将多伯努利混合部分近似为单个伯努利分量来处理, 即:
$ \arg \mathop {\min }\limits_f {D_{KL}}\left( {{f_{k|k}}\left( {{X_k}} \right)\parallel {{\hat f}_{k|k}}\left( {{X_k}} \right)} \right) $ | (17) |
其中,
$ D_{k|k}^u, {\left( {\hat r_{k|k}^i, \hat f_{k|k}^i} \right)_{i \in {\mathbb I}}} $ | (18) |
其中,
假设在
$ D_{k - 1}^u(X) = \sum\limits_{n = 1}^{{N^u}} {\omega _{k - 1}^{u, n}} GGIW\left( {X;\zeta _{k - 1}^{u, n}} \right) $ | (19) |
其中,
$ f_{k - 1}^i = GGIW\left( {X;\zeta _{k - 1}^i} \right) $ | (20) |
● 预测. 若有衍生事件发生, 则需要考虑原目标和衍生目标的预测, 所以预测阶段包括未发生衍生事件分量的预测和衍生分量的预测两部分.
$ \begin{split} D_{k|k - 1}^u\left( X \right) =& \sum\limits_{n = 1}^{{N^b}} {\omega _k^{b, n}} GGIW\left( {X;\zeta _k^{b, n}} \right) \\ & + \sum\limits_{n = 1}^{{N^u}} {\omega _{k|k - 1}^{u, n}} GGIW\left( {X;\zeta _{k|k - 1}^{u, n}} \right) \end{split} $ | (21) |
$ \omega _{k|k - 1}^{u, n} = \omega _{k - 1}^{u, n}{p_S} $ | (22) |
$ r_{NT, k|k - 1}^i = {p_S}r_{k - 1}^i $ | (23) |
$ f_{NT, k|k - 1}^i\left( X \right) = GGIW\left( {X;\zeta _{k|k - 1}^i} \right) $ | (24) |
其中,
$ \left\{ \begin{gathered} {\alpha _{k|k - 1}} = \frac{{{\alpha _{k - 1}}}}{\eta } \\ {\beta _{k|k - 1}} = \frac{{{\beta _{k - 1}}}}{\eta } \\ {m_{k|k - 1}} = F{m_{k - 1}} \\ {P_{k|k - 1}} = F{P_{k - 1}}{F^{\rm{T}}} + Q \\ {v_{k|k - 1}} = 2d + 2 + {{\rm{e}}^{ - {T_s}/\tau }}\left( {v - 2d - 2} \right) \\ {V_{k|k - 1}} = {{\rm{e}}^{ - {T_s}/\tau }}M{V_{k - 1}}{M^{\rm{T}}} \\ \end{gathered} \right. $ | (25) |
其中,
假设衍生事件发生在量测产生之间, 即在扩展目标跟踪滤波的预测步骤中, 且每个目标在每个时刻只产生一个衍生目标. 由于衍生目标相对于原目标的运动状态发生了变化, 若直接忽略对衍生目标的预测, 将会导致在衍生时刻无法检测跟踪到衍生目标, 对于该问题本文利用多假设模型对衍生目标建模, 在预测阶段根据给定的衍生假设值, 产生多个假设的衍生先验信息, 其中每个假设代表一个可能的衍生事件, 即假设衍生事件发生在
$ \xi _{t, k|k - 1}^{\left( {l, \kappa } \right)} = \left\{ {\gamma _{k|k - 1}^{(t)}, x_{k|k - 1}^{(t)}, X_{k|k - 1}^{(t)}} \right\}_{t = 1}^2, \; l = 1, 2, \cdots , d $ | (26) |
其中,
$ \zeta _{k|k - 1}^{\left( {1, l, \kappa } \right)} = \left( {{\alpha _{k|k - 1}}, \frac{{{\beta _{k|k - 1}}}}{\kappa }, m_{k|k - 1}^{\left( {1, l, \kappa } \right)}, P_{k|k - 1}^{\left( {1, l, \kappa } \right)}, v_{k|k - 1}^{\left( {l, \kappa } \right)}, \kappa V_{k|k - 1}^{\left( {l, \kappa } \right)}} \right) $ | (27) |
$ \begin{split} & \zeta _{k|k - 1}^{\left( {2, l, \kappa } \right)} = \\ & \left( {{\alpha _{k|k - 1}}, \frac{{{\beta _{k|k - 1}}}}{\kappa }, m_{k|k - 1}^{\left( {2, l, \kappa } \right)}, P_{k|k - 1}^{\left( {2, l, \kappa } \right)}, v_{k|k - 1}^{\left( {l, \kappa } \right)}, (1 - \kappa )V_{k|k - 1}^{\left( {l, \kappa } \right)}} \right) \end{split} $ | (28) |
其中,
$ m_{k|k - 1}^{\left( {1, l, k} \right)} = {m_{k|k - 1}} + \left( {1 - \kappa } \right)\sqrt {{e_l}} {H^{\rm{T}}}{v_l} $ | (29) |
$ m_{k|k - 1}^{\left( {2, l, \kappa } \right)} = {m_{k|k - 1}} - \kappa \sqrt {{e_l}} {H^{\rm{T}}}{v_l} $ | (30) |
$ P_{k|k - 1}^{\left( {i, l, \kappa } \right)} = {P_{k|k - 1}} , i = 1, 2 $ | (31) |
假设有
$ r_{T, k|k - 1}^i = {p_S}{\omega _{sp}}r_{k - 1}^i $ | (32) |
$ \bar f_T^i({X_{1:2, k, k - 1}}) \triangleq \bar f_T^i({X_{1, k|k - 1}}, {X_{2, k|k - 1}}) $ | (33) |
$ \begin{split} \bar f_T^i\left( {{X_{1:2, k|k - 1}}} \right) \approx &\sum\limits_{l = 1}^d {\delta \left( l \right)} GGIW\left( {X_{1, k|k - 1}^{\left( {l, \kappa } \right)};\zeta _{1, k|k - 1}^{\left( {l, \kappa } \right)}} \right) \\ &\times GGIW\left( {X_{2, k|k - 1}^{\left( {l, \kappa } \right)};\zeta _{2, k|k - 1}^{\left( {l, \kappa } \right)}} \right) \\ = &\sum\limits_{l = 1}^d {\delta \left( l \right)} \prod\limits_{t = 1}^2 {f_T^i} \left( {X_{t, k|k - 1}^{\left( {i, \kappa } \right)}} \right) \end{split} $ | (34) |
其中,
$ D_{k|k - 1}^u, {\left( {\hat r_{k|k - 1}^i, \hat f_{k|k - 1}^i} \right)_{i \in {\mathbb I}}} $ | (35) |
$ \left( {\hat r_{k|k - 1}^i, \hat f_{k|k - 1}^i} \right) = \left\{ {\left( {r_{NT, k|k - 1}^i, f_{NT, k|k - 1}^i} \right), \left( {r_{T, k|k - 1}^i, f_{T, k|k - 1}^i} \right)} \right\} $ | (36) |
● 更新. 假设给出如式(5)的预测PMB密度分布, 更新阶段的概率密度由以下4个部分组成.
(1) PPP漏检
$ \begin{split} & {D^u}\left( {X} \right) = \\ & \sum\nolimits_{n = 1}^{{N^u}} {\left( {\omega _1^{u, n}{GGIW}\left( {{X;}\zeta _1^{u, n}} \right) + \omega _2^{u, n}{GGIW}\left( {{X;}\zeta _2^{u, n}} \right)} \right)} \end{split} $ | (37) |
其中,
$ \omega _1^{u, n} = (1 - {p_D})\omega _{k|k - 1}^{u, n} $ | (38) |
$ \omega _2^{u, n} = {p_D}{\left( {\frac{{\beta _{k|k - 1}^{u, n}}}{{\beta _{k|k - 1}^{u, n} + 1}}} \right)^{\alpha _{k|k - 1}^{u, n}}}\omega _{k|k - 1}^{u, n} $ | (39) |
$ \zeta _1^{u, n} = \zeta _{k|k - 1}^{u, n} $ | (40) |
$ \zeta _2^{u, n} = \left\{ {\alpha _{k|k - 1}^{u, n}, \beta _{k|k - 1}^{u, n} + 1, m_{k|k - 1}^{u, n}, P_{k|k - 1}^{u, n}, v_{k|k - 1}^{u, n}, V_{k|k - 1}^{u, n}} \right\} $ | (41) |
(2) PPP更新
根据文献[17]的距离划分方法对量测集
$ r_C^u = \left\{ \begin{gathered} 1, \qquad\quad\;\;\;\; {\rm{if}} \; |C| > 1 \\ \frac{{{L_C}}}{{{\kappa ^C} + {L_C}}},\quad {\rm{if}} \; |C| = 1 \\ \end{gathered} \right. $ | (42) |
$ f_C^u\left( {X} \right) = \frac{{\displaystyle\sum\nolimits_{n = 1}^{{N^u}} {\omega _{k|k - 1}^{u, n}{p_D}L_C^{u, n}{GGIW}\left( {{X;}\zeta _C^{u, n}} \right)} }}{{\displaystyle\sum\limits_{n = 1}^{N{}^u} {\omega _{k|k - 1}^{u, n}{p_D}L_C^{u, n}} }} $ | (43) |
$ L_C^u = \sum\limits_{n = 1}^{{N^u}} {\omega _{k|k - 1}^{u, n}{p_D}l_C^{u, n}} $ | (44) |
其中,
(3) MBM漏检
$ r_\phi ^{j, i} = \frac{{r_{k|k - 1}^iq_D^i}}{{1 - r_{k|k - 1}^i + r_{k|k - 1}^iq_D^i}} $ | (45) |
$ f_\phi ^{j, i}\left( {X} \right) = \omega _1^i{GGIW}\left( {{X;}\zeta _1^i} \right) + \omega _2^i{GGIW}\left( {{X;}\zeta _2^i} \right) $ | (46) |
$ L_\phi ^{j, i} = 1 - r_{k|k - 1}^i + r_{k|k - 1}^iq_D^i $ | (47) |
其中,
$ q_D^i = 1 - p_D^i + p_D^i{\left( {\frac{{\beta _{k|k - 1}^i}}{{\beta _{k|k - 1}^i + 1}}} \right)^{\alpha _{k|k - 1}^i}} $ | (48) |
$ \omega _1^i = {\left( {q_D^i} \right)^{ - 1}}\left( {1 - {p_D}} \right) $ | (49) |
$ \omega _2^i = {\left( {q_D^i} \right)^{ - 1}}p_D^i{\left( {\frac{{\beta _{k|k - 1}^i}}{{\beta _{k|k - 1}^i + 1}}} \right)^{\alpha _{k|k - 1}^i}} $ | (50) |
$ \zeta _1^i = \zeta _{k|k - 1}^i $ | (51) |
$ \zeta _2^i = \left\{ {\alpha _{k|k{\text{ - }}1}^i, \beta _{k|k - 1}^i + 1, m_{k|k - 1}^i, P_{k|k - 1}^i, v_{k|k - 1}^i, V_{k|k - 1}^i} \right\} $ | (52) |
(4) MBM更新
$ r_C^{j, i} = 1 $ | (53) |
$ f_C^{j, i}\left( {X} \right) = {GGIW}\left( {{X;}\zeta _C^{j, i}} \right) $ | (54) |
$ L_C^{j, i} = r_{k|k - 1}^i{p_D}l_C^{j, i} $ | (55) |
PMB滤波器的更新步骤与PMBM滤波器类似, 详细的GGIW更新过程的参数计算参考文献[9].
4 实验仿真在区域
初始的两个目标及其各自的衍生目标参数设置如下:
$\left\{ {\begin{array}{l} {A_1} = 0.8, {a_1} = 1, {\gamma _1} = 10, {t_{b, 1}} = 1, {t_{d, 1}} = 100\\ {A_2} = 0.8, {a_2} = 0.5, {\gamma _2} = 5, {t_{b, 2}} = 51, {t_{d, 2}} = 90\\ {A_3} = 0.8, {a_3} = 1, {\gamma _3} = 10, {t_{b, 3}} = 4, {t_{d, 3}} = 100 \\ {A_4} = 0.8, {a_4} = 0.5, {\gamma _4} = 5, {t_{b, 4}} = 51, {t_{d, 4}} = 80 \end{array}} \right. $ |
每个扩展目标每个时刻产生的量测数量服从泊松分布,
假设场景中目标均做匀速直线运动, 目标的状态转移矩阵和噪声协方差分别为:
$ F = \left[ {\begin{array}{*{20}{c}} 1&1&0&0 \\ 0&1&0&0 \\ 0&0&1&0 \\ 0&0&0&1 \end{array}} \right] , Q = \left[ {\begin{array}{*{20}{c}} {\dfrac{{{T^4}}}{4}}&{\dfrac{{{T^3}}}{2}} \\ {\dfrac{{{T^3}}}{2}}&{{T^2}} \end{array}} \right] \cdot {0.1^2} \otimes \left[ {\begin{array}{*{20}{c}} 1&0 \\ 0&1 \end{array}} \right] $ |
为验证所提算法的有效性, 我们进行了50次蒙特卡洛实验, 同时选取广义最优子模式分配(GOSPA)距离[18]评估算法的性能, 其定义为:
$ \begin{gathered} d_p^{\left( {c, 2} \right)}\left( {X, \hat X} \right) = \\ {\left[ {\mathop {\min }\limits_{\theta \in {\Theta ^{\left( {|X|, |\hat X|} \right)}}} \sum\limits_{\left( {i, j} \right) \in \theta } {d_{GW}^{\left( c \right)}{{\left( {{x^i}, {{\hat x}^j}} \right)}^p} + \frac{{{c^p}}}{2}\left( {|X| - |\theta | + |\hat X| - |\theta |} \right)} } \right]^{\frac{1}{p}}} \\ \end{gathered} $ | (56) |
其中,
$ d_{GW}^{\left( c \right)}\left( {x, \hat x} \right) = \min \left( {c;{d_{GW}}\left( {{x^i}, {{\hat x}^j}} \right)} \right) $ | (57) |
其中,
图1(a)和图1(b)分别为标准的PMB算法和改进的PMB算法的跟踪轨迹. 通过对比图1两张图的跟踪轨迹可知, 衍生事件发生后, 改进的PMB算法利用多假设模型对衍生事件预测, 相比于标准的PMB算法能更有效地跟踪目标轨迹, 同时根据图2可知, 改进的PMB算法对目标数目的估计也更接近真实的目标数目.
GOSPA误差评估算法的整体性能, 图3显示了标准PMB算法和改进的PMB算法的GOSPA误差. 在扩展目标
为了验证改进的PMB算法在不同条件下的估计性能, 在其他仿真环境相同的条件下, 分别设置跟踪区域内的杂波率为
5 结论
针对标准PMB滤波器未对衍生扩展目标建模的问题, 本文利用多假设模型在PMB滤波的预测阶段对衍生事件进行预测, 利用随机矩阵法建模不同尺寸的椭圆外形, 使之更符合扩展目标衍生的假设. 仿真结果表明, 改进的算法能更有效地检测到衍生目标并进行跟踪, 但该算法对衍生的估计精度依赖于衍生假设值的选取, 因此如何合理地选取假设值提高衍生目标的估计精度还需要进一步研究.
[1] |
Bar-Shalom Y, Fortmann TE, Cable PG. Tracking and data association. The Journal of the Acoustical Society of America, 1990, 87(2): 918-919. DOI:10.1121/1.398863 |
[2] |
Granström K, Baum M, Reuter S. Extended object tracking: Introduction, overview and applications. arXiv:1604.00970, 2016.
|
[3] |
Mahler R. PHD filters for nonstandard targets, I: Extended targets. Proceedings of the 12th International Conference on Information Fusion. Seattle: IEEE, 2009. 915–921.
|
[4] |
Koch JW. Bayesian approach to extended object and cluster tracking using random matrices. IEEE Transactions on Aerospace and Electronic Systems, 2008, 44(3): 1024-1059. |
[5] |
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 |
[6] |
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 |
[7] |
裴佳. 基于随机超曲面的多扩展目标跟踪算法研究[硕士学位论文]. 西安: 西安电子科技大学, 2017.
|
[8] |
Chen YM, Liu WF, Wang XD. Multiple extended target tracking based on GLMB filter and Gibbs sampler. Proceedings of the 2017 International Conference on Control, Automation and Information Sciences (ICCAIS). Chiang Mai: IEEE, 2017. 26–31.
|
[9] |
Granström K, Fatemi M, Svensson L. Poisson multi-Bernoulli mixture conjugate prior for multiple extended target filtering. IEEE Transactions on Aerospace and Electronic Systems, 2020, 56(1): 208-225. DOI:10.1109/TAES.2019.2920220 |
[10] |
Granström K, Fatemi M, Svensson L. Gamma Gaussian inverse-Wishart Poisson multi-Bernoulli filter for extended target tracking. Proceedings of the 19th International Conference on Information Fusion (FUSION). Heidelberg: IEEE, 2016. 893–900.
|
[11] |
Xia YX, Granström K, Svensson L, et al. Extended target Poisson multi-Bernoulli filter. arXiv:1801.01353, 2018.
|
[12] |
Lian F, Han CZ, Liu WF, et al. Sequential Monte Carlo implementation and state extraction of the group probability hypothesis density filter for partly unresolvable group targets-tracking problem. IET Radar, Sonar & Navigation, 2010, 4(5): 685-702. |
[13] |
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 |
[14] |
苗露, 冯新喜, 迟珞珈. 基于GGIW-CPHD的衍生扩展目标跟踪算法. 计算机工程与应用, 2019, 55(9): 118-123. DOI:10.3778/j.issn.1002-8331.1801-0400 |
[15] |
马艳琴, 甘林海, 王刚. 基于δ-广义标签多贝努利的群分裂算法
. 现代雷达, 2018, 40(12): 46-51. |
[16] |
Cai RH, Xie Y, Wu SY, et al. Tracking the spawning of group with Poisson multi-Bernoulli mixture filter. Proceedings of the 2021 International Conference on Control, Automation and Information Sciences (ICCAIS). Xi’an: IEEE, 2021. 296–300.
|
[17] |
Granström K, Lundquist C, Orguner O. 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 |
[18] |
Rahmathullah AS, García-Fernández ÁF, Svensson L. Generalized optimal sub-pattern assignment metric. Proceedings of the 20th International Conference on Information Fusion (Fusion). Xi’an: IEEE, 2017. 1–8.
|
[19] |
Givens CR, Shortt RM. A class of Wasserstein metrics for probability distributions. Michigan Mathematical Journal, 1984, 31(2): 231-240. |