基于高阶正则化半监督学习的多跟踪器框架模型及融合策略研究

For arbitary target tracking in dynamically changing environments, the using condition limitations of the single tracker is usually regarded as the key point which prevents the improvement of target tracking algorithms. Thus, by mainly focusing on the problems exisiting in different single trackers such as low adaptiveness, volatile model update and etc., this project proposes a generic framework model composed with multiple trackers by extending the popular single trackers based on solitude model. A novel framework and a set of algorithms would be proposed by this project, and the goal of this investigation is to effectively resolve the most challenge problems of visual tracking in arbitary scenarios from a new viewpoint, and achieve more accurate and robust online target tracking by using unitary mathematical modeling strategy. The primary task that need be carried out in this project is about the common mechenisms of online learning sample selection and model learning by different trackers. The next important task is to do the research of multiple tracker framework model and fusion strategy based on high order regularization with semisupervised learning, which is the most important innovation of this project. With the initial achivements been accomplished by the first two tasks, we will continue with the study of optional status update and feedback function of the proposed tracking framework. The last essential task of the project is related to the solution of tracking error accumulation and re-target algorithms.Through the parallel simulation and implementation, we plan to improvement efficiency of our research achievments, which is promisingly to provide its contributions in different realistic applications such as video abstract, visual nevigation and visual sensor networks.

基于单分类模型的跟踪器适应性问题是制约复杂多变场景下任意目标跟踪效果提升的关键所在，而多跟踪器协同工作可以充分实现相异模型结构之间的优势互补从而平衡不同场景干扰之间的差异，降低分类误差，有效提高跟踪器的通用性。项目通过研究主流跟踪器所用生成模型和区别模型之间的共性机理，针对单一跟踪器存在的模型适应性差及全局参数更新不鲁棒问题，项目创新性地提出利用高阶正则化方法，构造基于多模型约束的多跟踪器复合框架模型，并通过半监督在线学习开展相关理论建模，参数优化和实验测试三方面的研究工作，验证所提出的框架模型既有对多类型跟踪器学习过程及结构化输出的融合能力，又有对真实环境中光照、遮挡、形变及相似背景等复杂应用场景的较强适应能力。进而研究统一框架下多模型参数的选择性更新及联合累积误差临界甄别问题，最终突破现有跟踪策略在复杂多变环境中的普适化限制，验证研究成果在视频摘要、视觉导航等方面应用的有效性及可靠性。

项目通过对相异跟踪器在线样本筛选与模型学习过程的共性机理与差异研究，在目标跟踪精度方面对不同跟踪器性能进行评估的基础上，提出多跟踪器目标状态匹配与融合的动态决策模型，突破在线目标跟踪中的多分类器融合、选择性更新与反馈、累积误差消除及目标重定位等关键技术，解决现有单一跟踪策略存在的模型适应性差，目标状态更新不鲁棒等主要问题。课题通过研究：相异跟踪器在线样本筛选与模型学习的共性机理研究；基于高阶正则化半监督学习的复合跟踪器结构化输出的状态决策模型；选择性状态更新策略与反馈机制研究以及复合跟踪框架下累积误差消除与重定位触发机制等内容，提出了一种基于高阶正则化半监督学习的多跟踪器框架模型，多跟踪器融合框架下的多模型参数选择性更新与反馈方法以及多跟踪器框架模型下的联合累积误差临界同步甄别方法将基于单模型/单跟踪器的跟踪框架模型泛化为普适的多模型/多跟踪器复合框架模型，在统一的数学模型下同时解决了具有结构性差异的多跟踪器高阶建模与状态融合问题，多跟踪器框架模型下的多模型参数的选择性更新问题和跟踪框架模型下的分类器累积误差临界值同步问题，充分发挥所提框架模型模型灵活性强、准确性高、鲁棒性好以及适应性广的特点，从新的角度解决目前制约复杂环境下任意形状物体在线跟踪的关键技术和难点问题，为监控与视频摘要、视觉导航、视觉传感网、体感与人机交互等实际应用提供了重要的基础理论支撑。