基于贝叶斯多核学习的多基地声纳信息预探测融合研究

To address the uncertain information from the large scale multistatic sonar network of low quality sensors, we intend to explore the predetection fusion within the framework of Bayesian multiple kernel learning for high quality contacts. Different from existing probability based methods, we intend to start with the geometric property of the multistatic sonar information, and regard the contact's measurement error covariance matrix as the kernel matrix. For the low quality of the contacts' meansurements, we intend to start from the regularization risk for arbitrary combination of kernels, and then take the correspondence between the regularization loss and the probability likelihood as a bridge, to establish a generalized Bayesian multiple kernel model for arbitrary combination of kernels, so as to fully explore the complex complementarity among the measurements. To address the uncertain information originated from multi-target and clutter, we intend to refer to the formulation of scale-mixed priors and the automatic clustering methods, and design the structured sparsity for the kernel weights, which can automatically discover the structure of the variables and allow negative correlation, so as to suppress the clutter as well as associate the contacts with their targets. For the large scale of the network, we intend to conduct the study on Bayesian inference within the framework of efficient variational approximation, so as to fuse the considerable contacts from the network with practical computational load. Finally, the proposed method will be applied to several typical trackers, so as to evaluate the effectiveness of this predetection fusion in multi-target tracking.

针对大规模低质量传感器组成的多基地声纳网络中的不确定信息，拟探索贝叶斯多核学习框架下捕捉层的预探测融合技术，以提供高质量的捕捉。有别于现有基于概率的方法，拟从多基地声纳信息的几何特性入手，将捕捉的测量误差协方差矩阵映射为核矩阵。对于低质量捕捉测量，拟从基于任意核组合形式的正则风险入手，以正则损失函数与概率似然分布间的对应关系为桥梁，建立任意核组合形式下的推广贝叶斯多核模型，以充分挖掘测量间复杂的互补关系；对于由多目标捕捉与杂波造成的信息不确定性，拟参考尺度混合先验的形式和自动聚类技术，为核组合权值设计能够自动发掘变量间分组结构，且允许负相关关系的结构稀疏先验，在抑制杂波的同时实现捕捉与相应目标的自动关联；对于大规模网络，拟在高效的变分近似框架下开展贝叶斯推理研究，以实用的计算负荷来融合大量的网络捕捉。最后，将其应用到多种典型的跟踪器上，评估该预探测融合技术在多目标跟踪中的效用。

针对大规模低质量传感器组成的多基地声纳网络中的不确定信息，研究了贝叶斯多核学习框架下捕捉层的预探测融合技术，以提供高质量的捕捉。有别于现有基于概率的方法，从多基地声纳信息的几何特性入手，将捕捉的测量误差协方差矩阵映射为核矩阵。对于低质量捕捉测量，从基于任意核组合形式的正则风险入手，以正则损失函数与概率似然分布间的对应关系为桥梁，建立了任意核组合形式下的推广贝叶斯多核模型，以充分挖掘测量间复杂的互补关系；对于由多目标捕捉与杂波造成的信息不确定性，参考尺度混合先验的形式和自动聚类技术，为核组合权值设计能够自动发掘变量间分组结构，且允许负相关关系的结构稀疏先验，在抑制杂波的同时实现捕捉与相应目标的自动关联；对于大规模网络，在高效的变分近似框架下开展贝叶斯推理研究，以实用的计算负荷来融合大量的网络捕捉。最后，将其应用到典型跟踪器上，评估了该预探测融合技术在多目标跟踪中的效用。