部件级表观模型的目标跟踪方法

Object Tracking Based on Component-Level Appearance Model

因受遮挡、运动模糊、剧烈形变等因素的影响,稳定且准确的目标跟踪是当前计算机视觉研究领域重要挑战之一.首先采用中层视觉线索的超像素描述目标/背景的部件,以部件颜色直方图作为其特征,并通过聚类部件库的特征集构建初始表观模型,部件表达的局部性和灵活性使该模型能够准确描述目标/背景;然后,利用贝叶斯滤波模型计算目标框的初始状态,并提出相似物体干扰的检测和处理算法以避免跟踪漂移,得到更健壮的结果;最后,为了减弱形变、遮挡、模糊对表观模型的影响以更好地保持目标特征,提出一种基于部件库的特征补集的在线表观模型更新算法,根据部件变化实时反映目标/背景的变化情况.在多个具有跟踪挑战的视频序列上的实验结果表明(共12个视频序列):与现有跟踪方法相比,该算法跟踪结果的中心误差更小,成功帧数更多,能够更准确并稳定、有效地跟踪目标物体.

Dealing with factors such as overlap, blurs from quickly moving and severe deformation, accurate and stable object tracking has become a critical challenge in compute vision field. First, in this paper, superpixels are used as middle level visual clue to describe the components of object/background with the color histograms of components as their features. The initial appearance model is proposed by clustering the features of a component library. The locality and flexibility of components representations allow the appearance model to describe object/background much more accurately. Then, the Bayesian filter model is used to compute the initial state of target region, and an algorithm is proposed to check and deal with the disturbance introduced by similar objects to avoid drift and obtain more robust tracking result. Finally, to reduce the influences of deformation, overlap and blurs to better preserve the features of object, an online appearance model update algorithm is developed based on the complementary set of the features of components library to enable the appearance model to reflect the real-time variation of object/background by the changes of components. Many experiments on video sequences with different tracking challenges （totally about 12 sequences） show that, compared with the existing object tracking methods, the proposed tracking algorithm results in less error of center position and more successful frame, and therefore can track an object more accurately, stably and effectively.