基于PHD滤波的检测前跟踪算法与重采祥技术研究

发布时间：2018-06-03 02:45

本文选题：概率假设密度 + 检测前跟踪　；参考：《浙江大学》2017年硕士论文

【摘要】：随着当代信息技术的飞速发展,无论是在军工层面还是在民用层面,多目标跟踪技术都拥有着广泛的应用前景。在目前复杂的电磁环境下,现有的传统跟踪算法对“低小慢”的微弱目标跟踪还存在一定的局限性,研究性能更好的多目标跟踪算法势在必行。本文在贝叶斯滤波的估计框架下,研究了当前较为热门的多目标跟踪算法——概率假设密度(Probability Hypothesis Density,PHD)滤波,并结合检测前跟踪(Track-Before-Detect,TBD)思想对微弱目标的处理优势,重点研究了基于PHD滤波的TBD算法理论及其具体实现。在传统的PHD-TBD滤波算法中,由于没有合理的新生粒子生成方案,以及没有满足使用PHD滤波的基本假设,严重限制了算法的滤波性能和实用价值。针对这些问题,本文研究一种改进的基于PHD滤波的检测前多目标跟踪算法,提出一种基于差分定位的自适应粒子生成方法,将新生粒子大量聚集在真实目标的位置周围,大大提高了新生粒子的有效性;同时建立新的观测模型,并通过引入相关阈值对观测数据进行预处理使得观测数据中的杂波数目近似服从泊松分布,令PHD滤波在TBD技术中可以更好的发挥其优势。仿真结果表明,所提出的算法可以提高目标数目估计的准确率,增强检测与跟踪性能,达到降低计算量的效果。在TBD算法中,每帧所要处理的数据量非常庞大,为了进一步提升算法的实时性能,重点研究了算法实现中的重采样过程,并提出了一种带有缓存机制的Metropolis Hasting(Buffered Metropolis Hasting,BMH)利于并行流水线操作的重采样方法,该方法在得到每个更新阶段的粒子和其对应权值时,就可以开始处理重采样步骤,不需要等待所有粒子的权值生成,并且能够保证权值较大的粒子被大量保存下来,维持算法的滤波跟踪性能。仿真结果表明,在相同的多目标仿真条件下,相比之前的方法,BMH采样可以获得较好的跟踪性能,并进一步提升了算法的实时性能。
[Abstract]:With the rapid development of modern information technology, multi-target tracking technology has a wide application prospect both in military industry and civilian level. In the current complex electromagnetic environment, the existing traditional tracking algorithms have some limitations on the weak target tracking of "low, small and slow", so it is imperative to study the multi-target tracking algorithm with better performance. In this paper, under the framework of Bayesian filtering, a popular multi-target tracking algorithm, probabilistic Hypothesis density (PHD) filtering, is studied, and the advantage of Track-Before-DetectTBD-based approach to weak targets is presented. The theory and implementation of TBD algorithm based on PHD filter are studied in detail. In the traditional PHD-TBD filtering algorithm, the filtering performance and practical value of the algorithm are seriously limited due to the lack of a reasonable new particle generation scheme and the failure to satisfy the basic assumptions of using PHD filter. In order to solve these problems, an improved pre-detection multi-target tracking algorithm based on PHD filter is studied, and an adaptive particle generation method based on differential localization is proposed. At the same time, a new observation model is established, and the number of clutter in the observed data is approximated by Poisson distribution by introducing the correlation threshold to preprocess the observed data. So that PHD filter in TBD technology can better play its advantages. The simulation results show that the proposed algorithm can improve the accuracy of target number estimation, enhance the detection and tracking performance, and achieve the effect of reducing the amount of computation. In the TBD algorithm, the amount of data to be processed in each frame is very large. In order to further improve the real-time performance of the algorithm, the resampling process in the implementation of the algorithm is studied. A resampling method with buffer mechanism for parallel pipeline operation is proposed. The method can process the resampling step when the particle in each update stage and its corresponding weight are obtained. It does not need to wait for the weight of all particles to be generated, and can ensure that the particles with large weights are saved in large numbers, and maintain the filtering and tracking performance of the algorithm. The simulation results show that under the same multi-objective simulation conditions, the BMH sampling method can obtain better tracking performance than the previous method, and further improve the real-time performance of the algorithm.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN713

【参考文献】