密集杂波条件下的多目标跟踪算法研究与实现

发布时间：2018-01-17 03:14

本文关键词：密集杂波条件下的多目标跟踪算法研究与实现　出处：《电子科技大学》2014年硕士论文　论文类型：学位论文

更多相关文章： 多目标跟踪 多模型 JPDA PHD JML-PDA

【摘要】：到目前为止,多目标跟踪算法可分为两大类:基于关联的多目标跟踪和基于随机有限集(Random Finite Set,RFS)的多目标跟踪,前者的代表算法为联合概率数据互联(Joint Probability Data Association,JPDA)算法;后者将的代表算法为概率假设密度滤波器(Probability Hypothesis Density Filter,PHDF)。它们之间互有优劣:基于关联的算法在密集杂波及信噪比较低的环境下能够获得较高的跟踪精度,但其中关联部分的计算量较大,PHD中没有关联部分,计算量较小,但它无法获得连续的航迹。此外,针对目标发生机动的情况,当前一般采用多模型(Multi Model,MM)或交互式多模型(Interacting Multi Model,IMM)滤波框架,对应衍生出了两类多目标跟踪算法的机动目标跟踪版本:IMM-JPDA与MM-PHDF。检测前跟踪(Track Before Detect,TBD)框架在杂波密度大的跟踪场景中能够获得较好的跟踪效果,到目前为止,基于RFS理论的多目标跟踪算法尚未得到完备的TBD模型,因此一般采用基于关联的跟踪算法:首先,通过多帧累积初始化目标轨迹,一般采用多目标轨迹起始算法:联合最大似然-概率数据关联(Joint Maximum Likelihood-Probability Data Association,JML-PDA)算法,在得到目标的初始状态及协方差矩阵后采用JPDA保持对目标轨迹的跟踪,称为结合的联合最大似然-概率数据关联(Combined JML-PDA and JPDA,CJML-PDA)算法。监测区域内的杂波密度和目标数量的增加会造成基于关联的多目标跟踪算法的组合爆炸问题,并且由于多帧累积算法往往需要用到多次观测数据,计算量很大,在串行运算的CPU平台处理器上无法实时实现,阻碍了它们的工程应用。自图形处理器(Graphic Processing Unit,GPU)被用于通用信号处理后,很多在传统架构上计算复杂度高的算法的实时实现成为了可能。本文针对多目标跟踪算法存在的一些缺陷,研究了多目标跟踪改进算法及其在硬件平台上的实现,主要的工作如下:(1)介绍了杂波条件下的多目标跟踪模型、基于关联的多目标跟踪算法JPDA及其机动目标跟踪版本IMM-JPDA;(2)介绍了RFS理论框架下的多目标跟踪模型、PHD滤波器、基于粒子滤波(Particle Filter,PF)实现的PHD滤波器以及PHD滤波器的机动目标跟踪版本:MM-PHD滤波和它的PF实现;(3)介绍了基于多帧累积的密集杂波条件下目标轨迹起始算法JML-PDA,基于JML-PDA算法与IMM-JPDA算法,结合两种算法的优势,提出了一种在密集杂波条件下机动多目标跟踪解决方案:CJML-IMM-PDA算法;(4)介绍了GPU的架构特性及其开发平台,提出基于GPU加速的CJML-IMM-PDA加速方案,并且基于仿真给出了其在GPU平台下相对于CPU的加速比,取得了良好的效果,为多目标航迹起始及跟踪算法的实时应用奠定了基础。
[Abstract]:Up to now, multi-target tracking algorithms can be divided into two categories: Multi-target tracking based on association and multi-target tracking based on random Finite set with random finite set. The representative algorithm of the former is Joint Probability Data Association JPDAA algorithm. The latter algorithm is represented by probability assumption density filter and probability Hypothesis Density Filter. There are advantages and disadvantages between them: the algorithm based on correlation can obtain high tracking accuracy in dense clutter and low SNR environment, but the computation of the correlation part is large. There is no correlation part in PHD, so the computation is small, but it can not get continuous track. In addition, multi-model Multi Model is generally used in the case of target maneuvering. MMM) or interactive multi-model Multi Model-IMM) filtering framework. Two maneuvering target tracking versions: IMM-JPDA and MM-PHDF.Track Before Detect are derived from two kinds of multi-target tracking algorithms. TBD framework can achieve good tracking effect in the tracking scene with high clutter density. Up to now, the multi-target tracking algorithm based on RFS theory has not yet got a complete TBD model. Therefore, the tracking algorithm based on association is generally adopted: firstly, the target trajectory is initialized by multi-frame accumulation. Multi-objective trajectory initiation algorithm is generally used: joint maximum likelihood probability data association (MMLP). Joint Maximum Likelihood-Probability Data Association. JML-PDAA algorithm uses JPDA to keep track of the target trajectory after obtaining the initial state and covariance matrix of the target. The combined JML-PDA and JPDA is called a combined maximum likelihood-probability data association. CJML-PDA algorithm. Monitoring the clutter density and the increase of the number of targets in the region will lead to the combination explosion problem of multi-target tracking algorithm based on association. And because the multi-frame cumulant algorithm often needs to use the multiple observation data, the computation is very big, and can not be realized in real time on the serial operation CPU platform processor. This hinders their engineering applications. Since graphic Processing GPUs have been used in general signal processing. Many algorithms with high computational complexity in traditional architecture are implemented in real time. This paper aims at the shortcomings of multi-target tracking algorithm. The improved multi-target tracking algorithm and its implementation on hardware platform are studied. The main work is as follows: 1) the multi-target tracking model under clutter is introduced. Multi-target tracking algorithm based on association JPDA and its maneuvering target tracking version IMM-JPDA; (2) this paper introduces a multi-target tracking model based on particle filter (Particle Filter), which is based on RFS theory. PHD filter and the maneuvering target tracking version of PHD filter: MM-PHD filter and its PF implementation; This paper introduces JML-PDA-based multi-frame cumulation-based dense clutter algorithm, and combines the advantages of the two algorithms based on JML-PDA algorithm and IMM-JPDA algorithm. A maneuvering multi-target tracking solution named: CJML-IMM-PDA algorithm under dense clutter is proposed. This paper introduces the architecture characteristics of GPU and its development platform, and puts forward a CJML-IMM-PDA acceleration scheme based on GPU acceleration. The speedup ratio compared with CPU on GPU platform is given based on simulation, and good results are obtained, which lays a foundation for the real-time application of multi-target track initiation and tracking algorithm.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN953

【共引文献】