【作者】 祁永庆；

【导师】 敬忠良；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 在多平台多传感器系统中,各传感器是互相独立工作的,其采样率也不全相同;测量数据的预处理是在各个传感器的局部坐标系里进行的。信息融合并不是多个传感器数据的简单的叠加,为了达到信息融合的目的,在数据相关之前需要进行时间和空间的配准,以形成时间和空间上统一的观测信息。传感器配准是信息融合的先决条件,也是提高系统整体性能的关键技术之一,它在武器系统、遥感、导航、智能交通、机器人和安全检查等军事领域和民用领域有着广泛的应用。本论文针对多平台多传感器配准算法的基础问题进行了深入、系统的研究。研究围绕着解决同类/异类传感器配准、动态偏差估计和异步传感器配准问题展开。主要研究成果如下:1.提出了高斯均值移动算法及其收敛的充分条件。首先研究了由核密度估计器推导均值移动算法的过程,以及均值移动算法的收敛性和轨迹平滑性。在此理论基础上,由极大似然概率密度函数推导出了高斯均值移动算法,给出并证明了高斯均值移动算法的收敛定理,该定理将当前的均值移动算法的收敛条件从一个凸核函数扩展为分段的凸核函数与凹核函数,拓宽了高斯均值移动算法的应用领域和应用空间,为后续提出的高斯均值移动传感器配准算法提供了理论基础。2.提出了基于信息融合的修正极大似然配准算法。针对传统双站被动定位系统的空间配准问题,分析了双站被动交叉定位法在有配准偏差的情况下存在导致盲区内目标不可观测的原因,结合多平台多传感器的项目背景与信息融合技术,提出了冗余信息补偿原理,用于修正极大似然配准算法。该算法利用系统中多站冗余信息,通过对盲区内的距离信息进行更新补偿,重新建立多站偏差补偿的配准模型,信息补偿后的目标位置被用于估计传感器偏差,实现了多平台被动传感器偏差补偿和目标定位。仿真结果表明,与现有方法相比,该算法可以有效提高无源定位系统的配准偏差的估计精度和系统的跟踪性能。3.提出了异类传感器的均值移动配准算法。针对多平台多传感器信息融合系统中的异类传感器配准问题,将极大似然法与均值移动法结合,提出了一种新的均值移动配准算法,并分析了算法的计算复杂度。该算法利用极大似然法估计目标状态,再将均值移动算法估计异类传感器的偏差。均值移动配准算法避免了极大似然配准算法中矩阵复杂的相乘和逆运算,减少了算法的计算量,且仿真结果显示,该算法较现有方法的偏差估计精度还略有提高。4.提出了用于动态偏差估计的高斯均值移动配准算法。现有的动态偏差配准方法是一种基于多帧多目标的方法,不能处理单一目标情况下的配准问题。针对多平台多传感器系统中动态变化的配准偏差问题,在高斯均值移动算法的理论研究的基础上,假设测量噪声服从高斯分布,推导出用于动态偏差估计的高斯均值移动配准算法。该算法的基本思想是首先利用扩展卡尔曼滤波器实时地估计目标状态,再利用高斯均值移动算法估计传感器的动态偏差。随后,又分析了该算法的计算复杂度,分析结果表明所提出的算法比现有方法的计算量小。利用该算法可以得到较好的配准效果,尤其是对于目标数较少的情况下,其配准效果较现有方法更加明显。5.提出了异步传感器的偏差估计算法。首先利用目标在同一时刻的真实目标状态对不同传输速率的传感器的测量值进行描述,再对不同传感器的测量值进行线性组合,目的是消除目标状态、构造关于传感器偏差的伪测量线性方程。然后,根据偏差的动态方程推导出了关于偏差的等效状态方程,最后利用卡尔曼滤波器对偏差进行估计。该算法可以有效地对异步传感器进行配准,与现有方法相比,它不受传输速率和传感器数量的限制,而且在偏差伪测量方程中没有引入自相关测量噪声。仿真结果表明,该算法与现有方法相比,不仅减小了计算量,而且提高了系统对偏差的估计精度。更多还原

【Abstract】 In multi-platform multi-sensor systems, each sensor is independent to work, and the data ratio is not completely same. The pretreatments of measurements are implemented in local coordinate of each sensor. Information fusion is not a process to accumulate the data from multiple sensors. In order to achieve the aim of information fusion, the temporal and spatial calibrations are performed to transform the measurements into uniform information in a common coordinate. Sensor registration is a precondition of information fusion, and is a key technique to improve the system performance. The sensor calibration has been applied in weapon systems, remote sensing, navigation, intelligent traffic, robot, security examination, other martial and civil areas.In this dissertation, multi-platform multi-sensor registration algorithms are studied. The research focuses on solving the problems of homologous / dissimilar sensor registration, dynamic bias estimation, and asynchronous sensor registration. The main contributions of this dissertation are summarized as follows:1. Gaussian mean shift and the sufficient condition for convergence are presented. The process that mean shift is derived from kernel density estimator, the convergence of mean shift, and its smooth trajectory property are studied firstly. Gaussian mean shift is derived from maximum likelihood density function, then the convergent theorem of Gaussian mean shift is proved. The theorem extends the current theorem from a convex kernel to a piece-wise convex and concave kernel, which widens the applied areas of Gaussian mean shift and provides the theoretic base for the latter chapters. 2. Modified maximum likelihood registration algorithm based on information fusion is proposed. Aiming at the space registration in two-observer passive tracking system, the reason that target is not observable in the blind spot is analyzed when there are registration biases in the system. The principle of redundant information compensation, based on the background of projects and information fusion, is proposed. The range information in the blind spot is updated by using the redundant information, and the registration model for bias compensation is rebuilt. The target location updated in the blind spot is used to estimate the sensor biases. The modified maximum likelihood registration algorithm achieves the aims of bias compensation and target location in multi-platform passive sensors system, which improves the estimation precision and the tracking performance of passive tracking system.3. Mean shift registration algorithm for dissimilar sensors is presented. The target state is calculated by maximum likelihood estimate, then mean shift is used to estimate the sensor biases. The computation complexity of the proposed algorithm is also analyzed. The algorithm is applied to estimate the biases of dissimilar sensors. The simulation results show that the new method improves the bias estimation precision, and furthermore, it reduces the complex computation.4. Gaussian mean shift registration algorithm for dynamic bias estimate is proposed. Based on the theoretic study on mean shift in the second chapter, Gaussian mean shift is derived with the assumption that the measurement noises are Gaussian. The basic idea of the proposed algorithm is that Gaussian mean shift combined with extended Kalman filter is implemented to estimate the dynamic biases. The algorithm complexity is also analyzed, and the computation of the proposed algorithm is less than other methods. The algorithm has better estimation performance than other methods, especially for that the number of targets is several.5. Asynchronous sensors registration algorithm is proposed. Firstly, the measurements from asynchronous sensors are expressed by the true target state in same time index. Then the linear function of the measurements is constructed to cancel the target state, and the pseudomeasurement equation of sensor biases is obtained. Thirdly, the equivalent bias state equation is derived from the bias dynamic equation. Finally, Kalman filter is used to estimate the biases of asynchronous sensors. The algorithm is effective to calibrate the asynchronous sensors without the limit of the data ratio and the number of sensors, and the autocorrelated measurement noises are not introduced in the pseudomeasurement equation. The proposed algorithm reduces the computation and improves the estimation precision.更多还原