【作者】 杨雷；

【导师】 赵春晖；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2007， 博士

【摘要】 随着多传感器复杂大系统的不断出现,多传感器数据融合技术已经受到广泛关注。目前,各国都在竞相投入大量的人力、财力进行研究,使得数据融合技术已经成为有效处理多源信息的一个非常活跃的领域。西方发达国家已经研制和开发了数个军用数据融合系统。我国数据融合的研究水平和国外发达国家相比还较为落后,尤其在复杂环境下,数据融合的一些相关技术还缺乏准确性和自适应性。本文主要研究了数据融合中三个最基本又最重要的问题:目标的检测、跟踪和识别关键技术的实现问题。对于多传感器多目标的检测识别环境,往往是复杂背景弱小目标的情况,多信息的关联和融合、跟踪都是以低信噪比、弱小目标的检测为前提的。系统的每个传感器对每个弱小目标的检测效果越好,其融合、跟踪能力越理想。因而本文在研究多传感器对多目标的融合跟踪之前,首先讨论了以红外弱小目标的检测为对象的复杂背景下低信噪比、弱小目标的检测方法。主要研究内容及创新工作如下:本文首先研究了红外图像的目标检测问题,针对低信噪比,海空背景的红外图像中的弱小目标检测问题,提出一种基于高斯内核的自适应滤波方法,并将它应用于红外图像的小波变换后的子图中抑制噪声,然后根据小波尺度分解的特点,把低频子图和水平子图以及垂直子图能量交叉加权,并进行小波重构。在此基础上,根据重构图像的信噪比,自适应的设定分割门限,分割红外图像中存在的可疑目标。实验表明,该算法易于工程实现,也能够满足实时性的要求。其次,对于多传感器多目标跟踪问题,数据关联是其中一项重要问题,也是实现多传感器数据融合的前提。研究了两种情况的数据关联-同类传感器和异类传感器。在同类传感器数据关联研究时,针对多维分配算法S≥3时求解复杂度随着问题规模的增大呈指数规模增大,提出把它转化为用粒子群优化算法求解的组合优化问题,并通过量测约束,更快更有效找到最优解,实现数据关联;在异类传感器数据关联时,提出同时利用雷达和ESM不同类型的信息的数据关联方法。该方法基于模糊C-均值算法,把动力学信息和来自ESM的特征信息有效结合,应用到聚类过程中,提高数据关联精度。再次,研究了目标航迹融合问题。采用一种利用模糊理论的自适应数据融合算法。对双传感器的滤波数据进行特征提取,并在一定的隶属函数和模糊规则下对其进行模糊推理,得到随目标机动情况自由调节加速度方差的系数调节值,使之保持对目标机动的快速响应,提高了跟踪精度。最后,在多传感器目标识别决策融合模型的基础上研究了基于模糊积分的目标识别决策融合的实现问题。对模糊积分应用于决策层目标识别的决策层融合的核心问题—模糊密度赋值,针对仅利用训练样本先验静态信息的获取模糊密度算法的不足,提出一种利用训练样本先验静态信息结合传感器识别信息集之间的支持可信度对模糊密度进行自适应赋值的方法。仿真实验结果表明,自适应动态模糊密度赋值方法通过传感器信息的支持度对传感器模糊密度进行实时的更新,减少了错误的信息对融合的影响,提高了多传感器系统的可靠性和鲁棒性。更多还原

【Abstract】 With the continuous appearances of complex multisensor large-scale systems, multisensor data fusion technique has attracted comprehensive attention. Studied with quantities of manpower and material resources in many countries, multisensor data fusion technique has been a very active field of processing multiple source information effectively. In western developed countries, some military data fusion systems have been researched and exploited. The study of data fusion in our country is still at a relatively low level, especially in complex environment, some correlative techniques in data fusion lack veracity and adaptiveness. Three important problems in date fusion, multisensor tagert detection, target tracking and recognition, are discussed in this dissertation. For multiple sensors multiple targets detection and identification environment, dim small targets are always in complex background, and the multiple information association, fusion, and tracking is on the premise of low SNR, dim small targets detection. The better results of dim small targets detection for each sensor of system, the better ability of fusion and tracking. Thus before multiple sensors multiple targets tracking and fusion, the detection methods for low SNR, dim small target in the complex background is researched first in this paper.The main research and innovation contents are as following:Firstly, target detection in infrared image is reseached. With the problem of the small target detection in low SNR and sky-sea background conditions, the paper proposed a method which based on the adaptive filter with Gauss function.This method is used in the sub-image which was obtained by wavelet analysis in the infrared image in order to restrain the noise. According to the trait of the wavelet analysis, we multiply the coefficient of the LL sub-image、HL sub-image and LH sub-image as the new LL sub-image, then use this coefficient by conversed wavelet transform. The results in the aspects of the mean, the standard deviation and the SNR of the image were better. Then, we found the threshold adaptively to segment the target in the image according to the SNR value. The simulation experiment results show the method is prone to realization in engineering and satisfies the meet of real time.Secondly, for multi-sensor multi-target tracking, data association is one of important problems, and is the precondition of data fusion. Two kinds of data association conditions - homogeneous sensors and heterogeneous sensors are reseaerched in this thesis. Consindered of data association for homogeneous sensors, the multiple dimensions assignment algorithm is used. This algorithm is NP-hard for S≥3, so it is transformed to a combination optimization problem and the Particle Swarm Optimization algorithm is used to solve this. Moreover, the measurements are restranited, then the solution can be found faster and more effective, and then data association problem is solved; Consindered of data association for heterogeneous sensors, a data association method of using two kinds of information from radar and ESM is proposed. It is based on the Fuzzy C-means algorithm, and dymatice information and characteristic information are used at the same time to cluster, so that the data association is enhanced.Thirdly, the problem of target track fusion was researched. A fuzzy theory based adaptive data fusion algorithm is used. The feature extraction was made for the filter data from two sensors and then the fuzzy illation was done under certain subordinate functions and rules, thereby a coefficient was got respinse to the practical conditions, and fusion results enhance the precision.Finally, the core problem in using fuzzy integral for decision fusion is to determine the fuzzy densities. According to shortage of the algorithm through that the fuzzy densities are determined only by aprior static information, a method of determining fuzzy densities adaptively is presented, which uses the apriori static information of the training samples and supporting reliability of each sensor’s information. The simulation experiment results show that due to real-time update of fuzzy densities through the method of determining fuzzy densities adaptively using supporting reliability of each sensor’s information, the effect of inaccurate information on fusion is depressed, and the reliability and robustness of multi-sensor system are enhanced.更多还原