未知量测噪声下随机集多扩展目标跟踪方法研究

发布时间：2018-03-27 18:55

本文选题：随机有限集　切入点：扩展目标　出处：《西安电子科技大学》2014年硕士论文

【摘要】：对于扩展目标来说,由于每个目标在每一个采样周期会产生多个量测,如若将量测和目标相关联,势必会存在巨大的困难,因此,研究一种更为实时有效的跟踪方法具有极其重要的现实意义和应用价值。近年来,由于计算复杂度比传统方法要小而且可以有效地处理传统跟踪算法中出现的某些问题,基于随机有限集的多目标跟踪方法受到了广泛的认同。本文针对基于随机有限集的跟踪方法展开重点研究,具体内容如下:1.基于高斯逆威舍特概率假设密度(GIW-PHD)的扩展目标跟踪算法。该算法在已知量测噪声协方差情况下,不仅考虑了目标的运动状态,而且考虑了目标的扩展状态。它将目标的运动状态建模为高斯分布,扩展状态建模为逆威舍特分布,通过量测数据来更新高斯分布以及逆威舍特分布中的参数,如自由度、逆尺度矩阵等,以此来达到跟踪目标的位置、大小、方向等信息的目的。2.基于随机超曲面概率假设密度(RHM-PHD)的扩展目标跟踪算法。该算法同GIW-PHD算法相似,在已知量测噪声协方差情况下并且考虑了目标的扩展状态。不过,RHM-PHD算法的量测建模方式与GIW-PHD算法有很大不同。RHM-PHD算法认为量测是由在目标表面随机分布的量测源再加上噪声所产生,GIW-PHD中的量测是由目标的运动状态再加上噪声所产生。此外,RHM-PHD算法是将表示目标扩展状态的参数嵌入到了运动状态矢量中,通过对运动状态矢量的更新来估计目标的形状、大小及方向。3.基于变分贝叶斯势均衡多目标多伯努利的扩展目标跟踪算法。该算法的优势在于它适用于量测噪声协方差未知的场景并且还提出了一种新的扩展目标的量测建模方式。核心思想在于对量测产生点状态和量测噪声协方差的联合概率密度用变分贝叶斯近似,之后再嵌入到势均衡多目标多伯努利框架中,在滤波更新得到量测产生点状态后,对其进行聚类从而得到扩展目标的估计状态。4.基于变分贝叶斯概率假设密度(VB-PHD)的扩展目标跟踪算法,同VB-CBMe MBer相同的是,该算法依然适用于量测噪声协方差未知的场景并且应用了VB-CBMe MBer中所提到的新的量测建模方式。然而,该算法是对量测产生点状态和量测噪声协方差的联合后验强度用变分贝叶斯近似,得到近似分布之后,对近似分布中的参数进行滤波更新后得到量测产生点状态,对其聚类从而得到扩展目标的状态估计。仿真结果表明,该算法与已知量测噪声协方差,且协方差为真实值时的CBMeMBer估计结果相当。
[Abstract]:For extended targets, since each target produces multiple measurements in each sampling period, if the measurement is associated with the target, there will be great difficulties. It is of great practical significance and practical value to study a more real-time and effective tracking method. In recent years, the computational complexity is smaller than the traditional method and it can effectively deal with some problems in the traditional tracking algorithm. The multi-target tracking method based on random finite set is widely accepted. In this paper, we focus on the tracking method based on random finite set. The specific contents are as follows: 1. An extended target tracking algorithm based on Gao Si inverse Weichet probability assumption density (GIW-PHD). The algorithm not only considers the moving state of the target, but also considers the moving state of the target when the noise covariance is known. Moreover, the extended state of the target is considered. It models the moving state of the target as Gao Si's distribution, the extended state as the inverse Weschet distribution, and the throughput data to update the Gao Si distribution and the parameters in the inverse Weschet distribution, such as degrees of freedom. The inverse scale matrix is used to track the position, size and direction of the target. 2. An extended target tracking algorithm based on the probability assumption density of random hypersurface (RHM-PHD) is proposed. The algorithm is similar to the GIW-PHD algorithm. When the measurement noise covariance is known and the extended state of the target is considered, the measurement modeling method of RHM-PHD algorithm is quite different from that of GIW-PHD algorithm. RHM-PHD algorithm thinks that measurement is recomposed by the random distribution of measurement sources on the surface of the target. The measurement in GIW-PHD caused by noise is caused by the moving state of the target plus noise. In addition, the RHM-PHD algorithm embeds the parameters representing the extended state of the target into the motion state vector. The shape of the target is estimated by updating the motion state vector, Size and direction. 3. An extended target tracking algorithm based on variational Bayesian potential equalization and multi-target multi-Bernoulli. The advantage of this algorithm is that it is suitable for measuring noise covariance unknown scenarios and proposes a new extended object. The key idea is to approximate the joint probability density of the measurement point state and the measurement noise covariance with variational Bayes. Then it is embedded into the multi-target multi-Bernoulli framework of potential equalization, and the state of the measurement generating point is obtained by the filter update. The estimated state of the extended target is obtained by clustering it. 4. The extended target tracking algorithm based on variational Bayesian probability assumption density (VB-PHD) is the same as VB-CBMe MBer. The algorithm is still applicable to the scene where the covariance of measurement noise is unknown and applies the new measurement modeling method mentioned in VB-CBMe MBer. However, This algorithm uses variational Bayes approximation for the joint posteriori strength of the measurement generating point state and the measurement noise covariance. After the approximate distribution is obtained, the parameters in the approximate distribution are filtered and updated to obtain the measurement generation point state. The simulation results show that the proposed algorithm is equivalent to the CBMeMBer estimation results when the measured noise covariance is known and the covariance is true.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.4

【相似文献】