基于不规则区域精确表征的非刚体运动目标在线跟踪研究

As the spring up of motion sensing, HD video coding and visual sensor networks, their further development raises the request for the support of accurate object tracking, which is a fundamental technique of these applications. Unfortunately, realistic object tracking is not easy because of the challenges from appearance extrinsic and intrinsic variations. Especially for a non-rigid object, since its geometry features may change totally different from its initial state as tracking goes by, it is not possible for the regular shape based online tracking-by-learning approaches to achieve a long-term accurate tracking. In this project, we innovatively propose a non-rigid object online tracking based on more accurate representation of irregular areas by using pair-wise gradient flow. By dynamic gradient change modeling, subsample representation with contex constraints and error denifition of subsample moving trajectory, some significant problems in non-rigid target tracking are supposed to be resolved. This project will carry out a comprehensive investigation from the aspects containing non-rigid object segmentation modeling, feature extraction and description for training samples, classifier online learning-by-classification mechanism, target border matting and tracking performance analysis. Based on the work of this research, a promising breakthrough for technical limitations is expected including accurate target segmentation, unit area projection and classification in the feature space, classifier error definition and p-n learning. And the final goal of this project is to resolve the challenge problems of inaccurate online training sample generation and weak update of t non-rigid target status during tracking. The potential value of this research project is to generalize the popular tracking methods based on regular representation to more generic and applicable non-rigid object tracking based on irregular representation. By using a novel integrated mathematic framework, this research project is supposed to overcome the key problems in the object tracking from a new point of view.

不规则区域的精确表征是非刚体目标跟踪的主要研究难点之一。本项目针对现有非刚体跟踪中在线产生训练样本误差大，目标状态更新不鲁棒的问题，将基于规则几何区域描述的跟踪泛化为普适的不规则区域跟踪，创新性的提出了一套基于不规则区域精确表征的非刚体目标在线跟踪方法，从新的角度解决目前制约任意形状物体在线跟踪的关键问题，在统一的数学模型下实现非刚体跟踪。研究内容主要包括：(1)构造基于目标动态边缘梯度动态变化的能量变化模型，实现非刚体目标精确分割；(2)研究基于临域约束的目标子样本特征空间映射方法，实现目标子区域的精确表征；(3)研究子样本轨迹参数化的分类器误差定义方法,实现目标检测在线正负约束学习；（4）实现目标边缘平滑及跟踪性能优化并对整体方法进行性能评估。本项目的研究在智能监控、人机交互、多媒体压缩、视觉传器网络等领域有重要研究意义和广泛应用前景。

不规则区域的精确表征是非刚体目标跟踪的主要研究难点之一。本项目针对现有非刚体跟踪中在线产生训练样本误差大，目标状态更新不鲁棒的问题，将基于规则几何区域描述的跟踪泛化为普适的不规则区域跟踪，创新性的提出了一套基于不规则区域精确表征的非刚体目标在线跟踪方法，从新的角度解决目前制约任意形状物体在线跟踪的关键问题，在统一的数学模型下实现非刚体跟踪。主要的研究成果包括：(1) 针对边缘梯度动态变化的非刚体精确分割建模问题，一个新颖的类吉布斯边缘梯度能量变化模型被提出，该模型的第一项通过基于图像透明度分布与任一帧图像像素信息的变化关系，同时利用帧内梯度和帧间的运动变化信息来体现了边缘梯度能量变化与平滑约束之间的关系，而通过一个迭代优化的过程以实现视频中连续帧之间非刚体目标区域的精确分割。达到精确分割非刚体目标不规则区域的目的；（2）针对目标训练样本的特征空间映射表征与约束关系建模问题，研究组通过创新性的增加该子区域在特征空间的相对目标中心的位置偏移量和子样本间的邻域约束为新的特征信息。利用一个基于狄利克莱和函数和潜在关联关系函数的复合分布函数来加以近似，达到将整体区域拆分为若干个子区域，实现在变换域空间对各子区域进行描述和表征目的；（3）针对子块目标检测在线正负约束学习与轨迹校准跟踪框架研究，一个全新的基于深度学习的在线学习模型被提出。该模型借助于深度CNN卷积网络对目标子样本集进行有效的特征提取新的样本学习结果也可进一步实现子样本轨迹的参数化并被定义为一系列约束条件，这使得原有跟踪框架中的正负样本学习策略的效能得到本质提升。而对于轨迹校准跟踪框架的研究也取得了显著成果；(4) 目标边缘平滑及跟踪性能的优化研究，，研究组成员通过结合图像形态学与图形学相关方法，利用目标边缘区域的各像素点所在的子块样本分类置信度实现目标边缘轮廓的概率密度表示，使得目标不规则区域的整体分割边缘的细化过程可以通过一个微分能量函数的最小化来加以实现，从而实现了非刚体目标分割后物体轮廓边缘的平滑处理。研究团队通过在多个国际通用视觉跟踪数据库上进行整体跟踪效果测试，进一步验证了所提出方法的有效性并发表一系列高水平国际会议和期刊论文。