【作者】 胡炜薇；

【导师】 杨莘元；

【作者基本信息】 哈尔滨工程大学 ， 通信与信息系统， 2007， 博士

【摘要】 随着现代跟踪环境显著变化，隐身与反隐身、对抗与反对抗措施，强机动、高杂波、低检测率和高虚警率等问题出现，使得跟踪系统设计遇到了强劲的挑战。而且为了得到观测对象更全面、更准确的信息，各种面向复杂应用背景的多传感器数据融合技术越来越受到人们关注，它必将是未来战争夺取胜利的关键。多传感器多目标跟踪是数据融合的底层关键技术。它将多个传感器信息有机合成，估计目标的运动状态，产生比单一传感器更优越的跟踪性能。本文以多传感器数据融合为背景，研究了多目标跟踪技术三方面的关键技术：机动目标状态估计、数据关联以及特殊情况处理。主要研究内容及创新工作如下：首先针对固定交互式多模型算法对目标进行状态估计时，对机动大的目标精度难以保证问题，提出基于期望系统噪声模型的自适应多模型算法。该算法自适应调整部分系统噪声模型，使之接近符合实际的系统噪声，提高了跟踪精度。其次，对复杂环境下多传感器多目标数据关联困难问题，提出两种数据关联算法，均提高了数据关联效果。一种是改进的基于模糊C-均值聚类算法的数据关联算法。传统的基于FCM算法的数据关联，在目标和杂波较密集时容易出现关联错误。而改进算法它采用粗、精关联相结合方法：先用粗关联排除部分杂波，再用模糊C-均值聚类来实现精关联。在精关联时把多传感器关联转化为多个单传感器关联，更有效的实现多传感器多目标数据关联。关联后对同类量测融合，达到很好的状态跟踪；另一种是基于粒子群优化算法的S-维分配算法。S-维分配算法的最大问题在于S≥3时它的求解复杂度随着问题规模的增大呈指数规模增大。对此将它作化为组合优化问题，给出用粒子群优化算法更快速求解的方法。先在S-维算法执行前，用跟踪波门确认有效量测，降低后续计算量。再把交叉算子和变异算子引入粒子群优化算法，在粒子群初始化步骤以及交叉、变异操作时充分考虑确认备选量测，缩小优化搜索范围，更快找到最优解实现量测关联。再次，针对传统的数据关联方法只利用那些与目标状态向量计算直接相关的运动学信息，用于关联的信息少、质量低，从而引起复杂环境下关联正确率低的现象，提出三种属性信息和运动信息结合的方案来提高关联性能。方案一：把分类信息和运动信息相结合，应用到综合概率关联算法。原综合概率关联算法，在回波密集时容易将各目标的综合量测混淆。因此利用分类混淆矩阵确定分类信息似然函数，再用它调节原算法只用运动学信息的似然函数，构造成一个综合两类信息的似然函数，从而修正互联概率，有效提高关联效果；方案二：同样利用分类信息似然函数，改造基于图模型航迹关联算法。针对原算法在复杂环境数据关联精度不高问题，同时利用两类信息修改顶点兼容函数和连线兼容函数，实现提高多传感器网络的数据关联目的；方案三：将多属性信息和运动学信息直接融合，应用到改进模糊C-均值聚类算法实现数据关联。在量测密集环境下，仅用运动学信息的聚类结果往往不是正确的关联结果，对此在所提算法中，在用运动信息同时，直接利用多特征信息计算聚类关键参数：样本间距离、隶属度，并能调节属性信息的影响度，从而更好的综合利用多属性信息实现数据关联。这种多信息融合使关联效果显著提高。最后研究了两类数据融合系统可能遇到的不确定情况：野值和非顺序量测。对于野值：当融合中心存在野值时，如果不作对应处理，融合数据会出现较大偏差。对此提出一种具有实时剔野能力的多传感器航迹融合方法。该方法利用模糊C-均值算法对多传感器点迹聚类，根据类内紧凑度及各点迹隶属度，识别并剔除航迹中的野值。最后用预测估计代替野值进行数据融合。该方法在实现数据关联同时有效快速解决野值问题；对于无序量测，传统滤波器无法直接处理。针对非线性系统中，扩展卡尔曼滤波无序滤波精度不高的问题，提出一种最优和次优的无序量测无迹卡尔曼(UKF)滤波器。它基于UKF滤波，重新推导了多步滞后量测转化为等价的一步滞后量测进行直接滤波估计过程，存储负担小。通过仿真比较，可看出所述滤波器能在非线性系统中，能更有效的利用无序量测并提高跟踪性能。之后又研究了多个同时滞后的无序量测情况的处理，提出用基于改进FCM聚类的数据关联算法对滞后量测进行数据关联，使所提的UKF无序滤波器能有效处理多个无序量测进行直接更新，提高跟踪精度。更多还原

【Abstract】 With the target tracking environment changing significantly, the tracking system design faces serious challenge of stealth and anti-stealth, ECM and anti-ECM, strong maneuvering, high clutter, low detection probability and high false-alarm ratio. The faced complex application background multi-sensor data fusion technology receives more and more attention, to get more complete and more accurate information of observation object. It is certain that the study of data fusion is the key of winning future war. The multi-sensor multi-target tracking is the bottom key technology of dada fusion. To get better estimate on the state of target motion than single sensor’s estimate, the multi-sensor information is synthesized organically. In the background of multi-sensor data fusion, three key technologies are discussed in this thesis focused on multi-sensor multi-target tracking, including maneuvering target state estimation, data association and special situation processing. The main research and innovation contents are as following:Firstly, when fixed Interacting Multiple Model (IMM) is used to track target with great maneuver, it is difficult to remain satisfactory performance. A new EVIM adaptive filtering algorithm based on expected system noise model is proposed. In this approach, a part of the system noise models are adjusted to adaptively match the unknown true mode, and track precision is raised.Secondly, according to the problem of difficult data association in complex environment, two data association algorithms are proposed; both of them improved performance of data association. The first is a modified Fuzzy C-means (FCM) cluster algorithm based data association algorithm. The original FCM based algorithm always makes mistake in data association in dense targets and clutter environment. Based on that, the proposed algorithm includes two parts: coarse and precision correlation: a part of interferences are eliminated during coarse correlation; then the remained sensors data are fuzzy clustered using FCM algorithm, through decomposing the data association problem of multi-sensor multi-target into that of several single sensor multi-target, the latter is of better performance for multi-sensor multi-target data association; finally, the better state estimation of the targets are gained after measure-to-measure fusion; Another proposed algorithm is the S-dimensional(S-D) assignment algorithm based on Particle Swarm Optimization algorithm. The main challenge in associating data from three or more scans of measurements, is that the resulting S-D assignment problem for S≥3 is NP-hard. Therefore, S-D assignment is formulated as a combinational optimal problem, and solved faster by Particle Swarm Optimization (PSO) algorithm. And before S-D algorithm carrying out, the track gate is used to confirm the validation of measurements, which reduces computation enormously. Moreover, the crossover operator and mutation operator are introduced to PSO, and the best solution for data association could be searched faster, because of reducing the search range, through the validated candidate measures are considered in particle swarm initiation, crossover rules and mutation rules.Thirdly, the traditional data association methods always use those direct correlation kinematics information, so the information for association is of little quantity and low quality, and consequently correct ratio of it is low in complex environment. For that, three methods are described in the following for combining the classification information and the kinematic information, to improve data association. The first method: the classification information and the kinematic information are combined, and both used to IPDA (Intergrated Probabilistic Data Association). For the original IPDA algorithm, the intergrated measurements of targets are always confounded in dense measurements environment. Thus the classification information likelihood function, which is defined by class confusion matrix, is used to adjust the kinematic measurement likelihood function in former algorithm. So the new synthetical likelihood function is structured on two kinds of information, and association probability is also modified. In this way, the classification information can aid data association effectively; the second method: Similarly the classification information likelihood function is used to modify graphical models based on track association algorithm. For original algorithm has low association precision in complex environment, the two kinds of information are both applied to rebuilding node compatibility functions and edge compatibility functions. The proposed algorithm improves data association performance in the sensor networks; the third method: the multi-character information and the kinematic information are fused directly, both used to Fuzzy C-means cluster based on data association algorithm. In dense measurements, when the kinematic information is only used to cluster, the association result is always not correct. So that in the proposed algorithm, the calculation of key parameter-distance and subjection functions is based on both kenematic information and multiple features of the target. Furthermore, the effect of features can be adjusted, thus data association is better realized based on the multiple features. The comprehensive information greatly raises the performance of data association.Finally, outlier and out-of-sequence measurement, two possible kinds of uncertain situation in data fusion system, are researched. For outlier, when it arrives in fusion center directly, it would induce large deviation to data fusion. So a method of real-time eliminating outlier multi-sensor data fusion is proposed. In this method, firstly the track of multi-sensor is clustered by FCM, and then the outliers are detected and eliminated, according to the compact degree and the membership. Finally forecasted estimation takes the place of outliers, and is sent to fusion. This method solves outlier problem rapidly and effectively, as well as data association problem; For out-of-sequence measurement (OOSM), traditional filters can hardly deal with it. According to extended Kalman filter (EKF) based OOSM filter has low precision for the nonlinear system, an optimal and a sub-optimal Unscented Kalman filter for out-of-sequence measurements are proposed. The procedure of out-of-sequence measurement updating, that converting multiple steps lag problem to one step lag problem, is deduced again based on UKF. And the method needs little memory burden. Simulation results show that the presented algorithms are more effective in utilizing out-of-sequence measurements and target tracking performance than those in EKF filter for nonlinear system. At last, the treatment of multiple OOSMs, which have the same lags, is researched. A FCM based data association algorithm is proposed to deal with multiple OOSMs. Thus the proposed UKF can update with multiple OOSMs, and raises track precision.更多还原