【作者】 龙建乾；

【导师】 付耀文；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2011， 硕士

【摘要】 目标跟踪技术一直是一个备受关注的研究热点,其在军事、民用领域都发挥着巨大的作用,有着重要的应用价值和广阔的发展前景。有限集统计学(Finite Set Statistics: FISST)理论在最近十多年开始应用于目标跟踪领域,由于其表现出来的对问题的直观描述性、理论推导的严谨性等特点,受到各国学者广泛关注和深入研究。针对该理论框架下最优多目标Bayes滤波器极高的复杂度问题,近年来先后有学者提出基于FISST理论的近似多目标跟踪算法,如概率假设密度(Probability Hypothesis Density: PHD)滤波、带有势分布的概率假设密度(Cardinalized Probability Hypothesis Density: CPHD)滤波等。尽管如此,概率假设密度类滤波器的运算复杂度在密集杂波背景下仍然过高;而另一方面,基于FISST理论的多目标跟踪算法通常未能给出前后时刻多目标状态估计的关联关系。本文在对PHD滤波、CPHD滤波进行深入研究的基础上,重点研究了密集杂波背景下PHD粒子滤波实现的改进算法、多模型CPHD滤波技术以及PHD多目标跟踪算法中的航迹形成等问题。具体研究内容如下:针对密集杂波背景下PHD滤波的粒子近似实现算法运算复杂度过高的缺点,借鉴传统跟踪算法中跟踪门限的思想提出了一种改进算法。该算法通过添加门限滤除杂波以降低滤波过程中的运算量。实验结果表明,在不影响滤波结果的情况下,改进的PHD滤波粒子近似实现算法提高了滤波效率。针对多机动目标的跟踪问题,本文提出了一种多模型扩展的CPHD滤波算法。该算法将传统的多模型方法与CPHD滤波相结合,在给出该算法详细的理论推导后,给出了该算法的粒子近似实现方法。实验结果表明,该算法对于多机动目标跟踪问题能够提供优越的性能。最后,本文针对基于FISST理论的多目标跟踪算法无法给出目标航迹的问题,提出了一种改进的通过添加粒子标签的航迹形成方法。与现有方法相比,该方法能够更好的处理存在衍生目标、目标交叉、杂波等复杂场景的航迹形成问题。实验结果验证了上述结论。更多还原

【Abstract】 Target tracking technique is a vital research topic for a long time, especially in the military and civilian areas. This technique has important application value and broad development prospects. Since the past ten years, Finite set statistics (FISST) theory has been beginning to be used for target tracking widely. Due to its demonstrated intuitive description to the problem and rigorous derivation, the FISST theory has attracted many research interests. Restricted to the computational intractability of the optimal Bayes filter in the FISST framework, some scholars have proposed many approximate algorithms based on the FISST theory for tracking multiple targets, such as the probability hypothesis density (PHD) filter, the Cardinalized probability hypothesis density (CPHD) filter, etc. Nevertheless, their computational complexity is still high in dense clutter environment. On the other hand, multi-target tracking algorithms based on the FISST theory always fail to give the track-value of the estimated targets.In this paper, based on the in-depth study on the PHD and CPHD filters, we give the improved algorithm for particle PHD filter in a dense cluttered environment, multiple model extension of the CPHD filter and improved tracks-value estimation algorithm for the PHD filter. The specific research contents are as follows: When the clutter is dense, the computational complexity of approximation particle algorithm for the PHD filter is too high. Inspired by the idea of threshold in traditional tracking algorithm, we propose an improved particle algorithm. The proposed algorithm reduces the amount of computation by adding threshold to filter the clutter. The experimental results show that, the improved particle PHD filter improves the filtering efficiency, without affecting the filtering results.Aiming at the tracking problem of multiple maneuvering targets, this paper presents a multiple model CPHD filter. The introduced filter combines the traditional multiple model method and the CPHD filter. After giving detailed theoretical derivation, we achieve the particle approximation method of the multiple model CPHD filter. Experimental results show that, the filter can provide superior performance for tracking multiple maneuvering targets.Finally, to account for the point that multi-target tracking algorithms based on the FISST theory can not give the tracks, we propose an improved track-value estimation method by adding the label to particles. Compared to the existing methods, the improved method treats complex scenes better, as spawning targets, target crossings and/or clutter. Experimental results validate the above conclusions.更多还原