【作者】 陈艳；

【导师】 万磊；

【作者基本信息】 哈尔滨工程大学 ， 船舶与海洋结构物设计制造， 2010， 硕士

【摘要】 目标跟踪技术被广泛的应用于监控、导航、避障等需要确定目标的数目、位置、运动速度和身份的系统中。尤其是水下目标跟踪技术,没有空中目标跟踪技术那么成熟,研究它具有理论意义和实际工程应用价值意义。本文主要研究基于高斯粒子滤波(GPF)的水下目标跟踪技术,主要分为两个部分:一部分是单目标跟踪,另外一部分是多目标跟踪。其中实现单目标跟踪是基础,而单目标跟踪的核心是选择合适的系统模型、观测模型和滤波算法。首先通过与光学图像进行比较分析,针对前视声纳图像含细节信息少的特点,建立了基于双特征匹配的观测模型,同时选择一阶自回归模型为状态转移模型。然后通过一维和二维的非线性非高斯运动模型的跟踪仿真实验证明:相对于著名的卡尔曼滤波算法来说,GPF不但可以解决线性高斯环境问题,还可应用于非线性非高斯问题;作为粒子滤波的改进,GPF不会出现粒子匮乏现象,不用进行重采样,相比之下,高斯粒子滤波更加简单易行,更适合应用于实际工程。最后序列图像实验和水池实验验证了本文所提出的单目标跟踪方法的有效性。在所提出的单目标跟踪的基础上,首先把GPF算法和最简单的数据关联算法——最近邻算法(NNDA)相结合,形成了GPF-NNDA多目标跟踪算法,但是它的跟踪效果并不理想,于是本文又提出了一种新的基于高斯粒子滤波和联合概率数据关联算法相结合的多目标跟踪算法,简称GPF-JPDA,最后通过声纳图像仿真实验和水池实验,比较了GPF-JPDA算法和其他算法的性能,实验证明本文所提出的基于前视声纳的GPF-JPDA多目标跟踪算法准确率高,实时性好,能够满足实际水下多目标跟踪的需要。更多还原

【Abstract】 Target tracking technology is widely used in surveillance, navigation, obstacle avoidance and so on, which needs to make sure the targets’ number, locations, speeds and identities. Especially the underwater target tracking technology is not as mature as air target tracking technology, research of which has very important significance.This paper is concerned with the study on underwater target tracking in sonar images using Gaussian particle filter (GPF), and it is divided into two parts: the one is single target tracking, and the other is multi-target tracking. And the single target tracking is the basis of multi-target tracking. The most important step of single target tracking is to select an appropriate target dynamics model, a proper measurement model and a good filter. As the forward-looking sonar images are with less detail information of the targets like contour and color compared with the optical image, the first-order autoregressive process equation is selected as state transition model and the weight of a particle is evaluated according to matching its two characteristics of moment invariant and area with the corresponding characteristics of the target. simulation results of one-dimensional and two-dimensional non-linear non-Gaussian tracking model show that the Gaussian particle filter can not only solve the linear Gaussian problem, but also can be applied to non-linear non-Gaussian problems , comparing with the well-known Kalman filter who is only suitable for the linear cases; and as a improvement of particle filter it eliminates particle impoverishment without resampling, therefore it is easier to practice and more suitable for solving the practical engineering problems. Tank experiments are carried out .Results demonstrate the method’s advantages which is showed in the simulation. Based on the single-target tracking method, firstly GPF is joined with data association of nearest neighbor data association(NNDA), which is the easiest multi-target tracking method named GPF-NNDA, but it can’t fulfill the task of tracking multiple closed targets. Then a new multi-target tracking method is proposed in this paper which is based on the Gaussian particle filter and joint probabilistic data association named GPF-JPDA. Simulation and tank experiments of multi-target tracking are performed using GPF-JPDA and some other data association to test the performance of the presented method in this paper which implies that GPF-JPDA has the advantages of good robustness, high accuracy and real-time characteristic, and it is efficient in underwater multi-target tracking based on sonar images.更多还原