【作者】 申屠晗；

【导师】 薛安克；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2014， 博士

【摘要】 以信息融合理论为基础的目标跟踪技术指利用雷达、红外、声纳等多种传感器的观测信号对未知目标的数量、位置、速度、身份等状态进行估计的过程。目标跟踪技术已被广泛和成功地应用于军事和民用领域,包括国土防空、导航、制导、探测、定位、交通、制造、金融和医疗等。近年来,目标跟踪环境日趋复杂,强杂波干扰、大观测误差、低信噪比和高机动等大量不确定因素使传统目标跟踪技术受到严重挑战。由于传统目标跟踪技术采用单向开环融合模式,不利于信息的存储、挖掘与复用,导致复杂环境下跟踪效果下降。为此,本文将信息反馈融合的理念融入传统目标跟踪技术,较系统地研究了面向目标跟踪问题的信息反馈融合框架、方法和工程算法,主要研究成果如下：1)针对传统信息融合方法缺乏信息反馈融合机制的缺陷,首创性地提出了空域信息融合平面与时域信息融合空间的概念,并构建了一般性的信息反馈融合框架。2)针对传统变结构多模型方法缺乏后验信息反馈融合机制的缺陷,提出信息反馈融合最小熵变结构多模型方法(MEVSMM, Minimum Entropy Variable Structure Multiple-Model)和次优算法,将后验香农信息熵作为评价模型序列集的质量指标,利用信息反馈融合完成MSA(Model Sequence Set Adaptation)优化过程。与传统多模型方法相比,本方法估计精度更高、鲁棒性更好。3)针对复杂观测误差环境下传统MSE(Minimum Shannon Entropy)方法跟踪效果退化的缺点,提出用GE(Geometrical Entropy)度量来补偿SE(Shannon Entropy)度量的估计偏差,并构建了信息反馈融合最小几何熵多模型方法(MGEMM, Minimum Geometrical Entropy Multiple-Model)和两个次优算法。与传统MSE方法相比,在先验观测误差分布与实际分布不一致时,本方法所得到的跟踪估计结果精度更高、鲁棒性更好。4)针对传统多模型融合方法缺乏历史信息反馈处理机制的缺陷,提出历史信息反馈融合多模型估计方法(HFMM, Historical Feedback MM)和次优算法来解决历史估计信息的反馈融合问题。与传统多模型方法相比,本方法提升了历史融合信息的利用率,从而获得精度更高、鲁棒性更好的融合估计结果。5)针对传统PHD(Probability Hypothesis Density)跟踪方法在缺乏目标初生先验信息时失效的缺陷,提出历史信息驱动反馈融合多目标跟踪方法HIFMTT (Historical Information Feedback Fusion Multiple Target Tracker)和次优算法。与传统PHD方法相比,无论是否具备目标初生先验信息,本方法都能较好地完成多目标跟踪工作。6)针对传统HIFMTT方法在跟踪隐身目标时结果退化的缺陷,提出信息预测反馈融合多目标跟踪方法IPFMTT (Information Prediction Feedback Fusion Multiple Target Tracker)和次优算法来克服隐身目标观测检测率较低的困难。与传统HIFMTT方法相比,本方法以小幅增大跟踪虚警率的代价大幅提升了跟踪检测率。更多还原

【Abstract】 The target tracking technology based on information fusion theory refers to the process of real-time estimating the states of the target quantity, position, velocity, identity, etc., based on the observations from the multiple sensors, such as radar, infrared, sonar, etc. The target tracking technology has been applied broadly and successfully to the military and civilian applications, including the national defense, navigation, guidance, detection, localization, transportation, manufacturing, finance, medicine, etc. The target tracking environments become more complex recently and the traditional target tracking technologies are challenged seriously by handling high level of uncertainties from environment, such as dense clutters, big observation errors, low SNR, high maneuvers, etc. Most traditional target tracking technologies adopt the single direction and open loop fusion modes and are weak in accumulating, mining and reusing the information. In consequence, the fusion performance of the traditional target tracking technologies will degenerate in complex environments. To this end, this thesis studies the information feedback fusion framework, methods and algorithms for the target tracking problems systematically, and the main results are as follows,1) The traditional information fusion methods have the drawbacks of lacking of the information feedback mechanism. To this end, we first propose the concepts of the spacial information fusion plane and the temporal information fusing space, then construct a general information feedback fusion framework.2) For the traditional variable structure multiple model methods have the drawbacks of lacking of the posterior information feedback mechanism, we propose the information feedback fusion minimum entropy variable structure multiple-model method (MEVSMM) and the sub-optimal algorithm. The proposed method takes the posterior Shannon Entropy as the optimization criterion for the model sequence set, and accomplishes the MSA process using the information feedback fusion mechanism. Compared to some traditional MM methods, the proposed method achieves more robust and accurate estimating results.3) The traditional MSE based algorithms will degenerate in some complex observation error situations. To this end, we set up the GE measure to compensate the ill estimation problem deduced by the SE measure. Further, we propose the information feedback fusion minimum geometrical entropy multiple-model method (MGEMM) and two sub-optimal algorithms. Compared to some traditional MSE methods, the proposed method achieves more robust and accurate results when the prior observation error distribution is inconsistent with the real situation.4) The traditional multiple-model fusion methods have the drawbacks of lacking of the mechanism of historical information feedback processing mechanism. To this end, we propose the historical feedback fusion multiple model method (HFMM) and the sub-optimal algorithm to realize the feedback fusing of the historical information. Compared to some traditional MM methods, the proposed method improves the information utilization ratio and achieves more robust and accurate estimating results.5) For the traditional PHD tracker will lose its efficacy when there is little new target prior information, we propose the historical information feedback fusion multiple targets tracker (HIFMTT) and the sub-optimal algorithm. Compared to the traditional PHD methods, no matter the newborn target prior information is available or not, the proposed method can achieve desirable multiple-target tracking results6) The traditional HIFMTT will deteriorate when tracking the stealth targets. To this end, we propose the information prediction feedback fusion multiple targets tracker (IPFMTT) and the sub-optimal algorithm to handle of low observation detection rate of the stealth target. Compared to the traditional HIFMTT methods, the proposed method improves the tracking detection rate significantly with only a slightly higher false alarm rate.更多还原