【作者】 吴立珍；

【导师】 沈林成；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2013， 博士

【摘要】 对战场空间内高价值目标的感知是无人机（Unmanned Aerial Vehicle，UAV）系统实现自主控制的基础和关键。基于图像特征的目标建模作为目标信息处理的基本环节，在UAV战场感知流程中起着承上启下的作用。然而，由于复杂战场环境中UAV实际获取的目标图像通常存在较大的不确定性，使得建立可靠的目标表示与描述成为一项非常具有挑战性的工作。针对这一问题，论文以UAV战场感知为研究背景，结合目标特征匹配、目标类检测以及目标跟踪三个典型应用开展目标特征建模方法的研究，主要工作及创新点如下：1）在深入分析和讨论融合上下文信息（contextual information）的目标特征建模理论的基础上，总结提出了三种目标特征建模框架。针对目标特征建模中的特征提取和表达问题，详细比较了不同的局部不变性特征检测与描述方法。同时，重点分析了具有一体化上下文建模优势的条件随机场（Conditional Random Field，CRF）模型，研究了不同扩展形式CRF模型的上下文信息利用能力以及面向目标特征建模应用的关键问题。2）提出了基于局部特征分组和多目标优化的目标特征匹配方法，通过对特征上下文的利用能够有效克服图像模糊引起的误匹配。首先，使用图像仿射不变区域特征对不变性点特征做出范围约束，并引入特征间的几何定序信息实现了局部特征分组设计；其次，完成了特征匹配的多目标优化问题建模，提出一种新的离散多目标粒子群优化算法实现了局部特征分组的最佳匹配搜索。同不考虑上下文信息的局部特征匹配方法相比，该方法有效提高了目标特征匹配的鲁棒性。3）提出了一种基于分层上下文表达的扩展CRF模型，实现了显著类内变化条件下单实例目标类的有效检测。针对经典CRF模型在目标特征建模中的不足，基于部件建模理论提出了CHCRF（Contextual Hierarchical Part-Driven CRF）模型，详细设计了模型的图结构、势函数、以及参数估计与模型推断方法。同时，提出了该模型的加权邻域结构形式，解决了邻域节点间依赖程度的表达问题。CHCRF模型的主要优势是能够同时利用目标图像中潜在的部件层上下文和观测层上下文信息，实现目标类及其内部隐含信息的统一建模。4）针对UAV多实例目标检测问题，提出了基于多任务分解的MFCRF（Multi-task Factorial CRF）模型，解决了模型设计和实现中的关键问题。同单一标记任务的CRF模型相比，MFCRF模型的主要优势是建立了目标多元部件标记任务和二元目标标记任务的融合框架，能够以互补的形式融入不同任务标记间的相关性信息，从而更好地利用标记层上下文。实验结果表明，相比于现有的其它扩展CRF模型，MFCRF模型具有更强的目标建模与检测能力。5）提出了基于时空上下文CRF模型的目标跟踪算法，有效解决了复杂动态场景中视觉目标跟踪的模型漂移问题。通过将经典CRF模型扩展到时域空间，提出了融入时空上下文信息的三维STCRF（Spatial-Temporal CRF）模型，设计了模型的时域邻域系统。该模型不仅能够借助目标标记的空间邻域相容性应对目标姿态变化和图像模糊，而且可以引入时域平滑约束对目标标记在跟踪中的连续性建模。进一步，设计了基于STCRF模型标记结果的目标跟踪算法和模型更新策略，实现了目标的稳健描述与跟踪。实验结果表明，论文所提算法相比于传统的局部特征匹配跟踪，平均跟踪误差显著降低。更多还原

【Abstract】 The perception of high-value targets in battle space is a foundational and keyproblem for an unmanned aerial vehicle (UAV) system to achieve autonomous control.As one of the essential contents of object information processing, object featuremodeling plays a connecting role in the process of UAV battlefield awareness. However,due to the high complexity and dynamics of the battlefield environment, the actualimages acquired by UAV usually have large uncertainty, which makes the establishmentof a reliable object model become a very challenging task. Under such a background,this dissertation focuses on the problem of object feature modeling, including threetypical applications, i.e. object feature matching, object category detection, and objecttracking. The main work and contributions of this dissertation are summarized asfollows:1) On the basis of in-depth analysis and discussion of the object feature modelingtheory with contextual information, the dissertation proposes three object featuremodeling frameworks. Aiming at the problems of feature extraction and expression inobject modeling, different detection methods and description types of local invariantfeatures are discussed in detail. Meanwhile, conditional random field (CRF) model,which has the unified framework for contextual information modeling, is analyzedemphatically. The abilities of using contextual information and key problems in objectfeature modeling for various extended forms of CRF model are studied.2) A novel feature point matching method that benefits from local-feature-groups(LFGs) and the multi-objective optimization theory is proposed, which can effectivelyreduce the mismatches caused by image ambiguities. Fisrt of all, LFGs are constructedthrough grouping affine invariant point features and regional features. Regional featuresmake a range constraint to point features, and geometrical ordering contextualinformation in each group is introduced to enhance robustness and discrimination ofLFGs. Then, feature point matching is formulated as a multi-objective optimizationproblem, and the optimal search is realized by using a novel discrete multi-objectiveparticle swarm optimization algorithm. Experimental results show that the proposedmethod can effectively improve robustness of object feature matching while comparedwith those without contextual information.3) In order to overcome the drawbacks of the classical CRF model for objectcategory detection, a contextual hierarchical part-driven CRF (CHCRF) model isproposed in the dissertation, which can effectively detect single-instance object categorywith notable intra-class variations. The graph structure, potential functions, parameterestimation and inference methods of CHCRF model are designed, respectively.Meanwhile, a weighted neighborhood structure is developed to capture the degree of correlation between connected nodes in the model. By using a two-layer hierarchicalformulation of labels, CHCRF model can effectively represent label-level context andobservation-level context simultaneously, and achieve unified modeling of objectcategory and its internal latent context.4) For multi-instance object detection, the dissertation proposes a multi-taskfactorial CRF (MFCRF) model to tightly combine the task of multiclass part labelingwith the task of binary object labeling. By fusing different types of labels, the MFCRFmodel can make full use of label-level contextual information. Experimental resultsshow that the modeling capability and detection performance of MFCRF model aresuperior to other existing extended forms of CRF model.5) To solve the drift problem of object model for visual tracking in complexdynamic scenes, a novel CRF model with spatial-temporal contextual information isproposed. By extending the classic CRF model in the time domain, the definition of thetime-domain neighborhood system is developed, and a spatial-temporal CRF (STCRF)model is proposed. The main characteristic of the model is that it can not only use thecompatibility of spatial neighborhood to deal with pose variations and image ambiguity,but also use the time domain smoothness constraint to model the continuity of objectlabels. Furthermore, an object tracking algorithm and a model update strategy based onlabeling results from STCRF model are designed to achieve robust object descriptionand tracking. Experimental results show that the proposed algorithm can significantlyreduce the average tracking error while compared with traditional feature matchingtracking methods.更多还原