基于在线距离度量学习的自适应视觉跟踪方法

Appearance modeling and visual matching are two key issues for the success of a visual tracker. However, most of existing target tracking methods predefine the above two parts and consider them separately, which may give poor performance, especially for significant appearance variations. Distance Metric Learning (DML) theory provides an opportunity for solving this problem. In order to realize the unification of appearance modeling and visual matching for long-term tracking, this project aims at the systematic research on the visual tracking mechanism and method under online DML framework. Firstly, for an unknown object without any prior information of its appearance, this project will propose an object and background modeling mechanism based on feature saliency prediction and dimensionality reduction, which reveals feature descriptors’ ability for target characteristics and achieves efficient target/background descriptions in virtue of prediction function. Secondly, this project will seeks an appropriate metric in the feature space and propose a metric learning based structural appearance model for more accurate matching of different appearance. An online local distance metric learning algorithm is proposed that learns a discriminative and adaptive metric, thereby better distinguishing the visual object of interest from the background. By taking the data local distribution and target motion into consideration, a continuous closed-form output will be provided efficiently. Meanwhile, in order to deal with the problems of appearance variations, occlusions and drift, an adaptive template library updating strategy will be proposed based on tracking state judgement and a closed-loop feedback mechanism. Through the research of this project, an adaptive visual tracking technical framework based on online DML will be constructed, which can provide a practical and feasible solution to long-term robust tracking in complex dynamic environments; Moreover, the achievements will also expand DML theory, and offer reference for other areas employing DML theory.

外观建模和特征匹配是影响视觉跟踪成败的两个关键环节。现有方法通常“预先指定”且“割裂考虑”二者，导致跟踪性能下降。距离度量学习(DML, Distance Metric Learning）理论为解决这一问题提供了契机。本项目拟对在线DML框架下的视觉跟踪机理和方法进行系统研究，以实现快速长程持续跟踪。具体包括：（1）研究基于特征显著性预判和压缩降维的外观建模，揭示特征描述子对目标自身特性的刻画能力，建立预判函数与特征描述子之间的对应关系；（2）研究匹配度量的局部在线增量学习方法，分析局部数据空间分布和目标运动特性对度量更新的影响，推导数据驱动的高效连续解析解；（3）研究跟踪状态判定和闭环反馈机制，建立自适应的模板库更新模型，有效克服跟踪漂移。通过本项目的研究，可以为复杂动态场景下的鲁棒视觉跟踪提出切实可行的解决方案；同时，取得的成果也将充实扩展DML理论本身，为其他领域的应用提供参考。

在视觉跟踪实际应用场景中，可获取的目标先验知识通常很少，这对于传统的基于预定义距离度量的跟踪算法而言是很大的挑战。而且，预定义的度量难以适应目标和环境的变化，易导致跟踪失败。因此，本项目研究了具有在线学习能力的视觉跟踪算法，构造随跟踪任务的推进而不断更新的度量矩阵，提高了算法的适应性和精度。.本文首先提出了结合距离度量学习的鲁棒在线视觉跟踪算法。该算法将视觉跟踪任务视为前、背景的二分类问题，并随着视频推进不断通过信息论手段在线更新基于欧式距离的分类器，从而实现稳定的跟踪。为了减少跟踪漂移，设计了对目标变化具有适应更新能力的快速更新模板和用于降低模板污染的稳健更新模板。为了提高速度，提出基于随机主成分分析降维的密集尺度不变特征描述，以在尽量保证高维特征优良属性的前提下尽可能降低特征维度、减小运算负担。.基于相关滤波器的跟踪算法定位精度高、速度快，但是它只能实现以像素为单位的低精度离散定位，易造成的跟踪漂移。同时其尺度估计模块精度严重依赖于目标定位的准确性。针对这些问题，本项目提出一种基于连续相关滤波器的亚像素跟踪算法。在目标定位模块中，算法利用卷积神经网络的高判别力特性，通过构建的卷积操作子实现对目标的连续定位，提高了算法的整体定位精度。在尺度估计模块，通过构建基于HOG特征的多金字塔目标尺度追踪器，实现了微小定位误差的修正及鲁棒的目标尺度估计，提升了算法整体性能。.在长程视频中，当目标被严重遮挡时跟踪器容易丢失目标，这是跟踪的难点问题。针对这一难点，本项目提出了基于Random-Ferns和基于Mean-Shift理论的目标重追踪方法。算法在视频目标的稳定帧建立目标重追踪模型，当判断目标遮挡消失后，算法在目标消失位置使用大范围的搜索，当目标再度出现时以重新获取丢失目标。.在OTB-100和VOT2015两个权威视频数据集上，将提出的算法与现今公认优秀的算法分别进行测试，实验结果表明提出的算法能够很好地应对不同场景下的跟踪任务，具有很强的竞争力。