【作者】 王军；

【导师】 王岳环；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2016， 博士

【摘要】 视觉目标跟踪(Visual object tracking, VOT)技术是计算机视觉的一个基础和关键的研究方向,近年来一直是学术界和产业界关注的热点之一。尽管近年来国内外研究者在目标跟踪上取得显著的进步,但是如光照变化、形变、遮挡等复杂目标外观变化的干扰以及跟踪器难以从跟踪失败中恢复的难题,使得现有跟踪算法很难实现目标长时持续跟踪。此外,目标发现阶段运动阴影检测算法的效率也会影响视频监控应用中跟踪器初始化的精度,引发跟踪失败。本文针对这些问题开展了以下研究：首先,为提升跟踪算法应对目标复杂剧烈外观变化的能力,本文研究了一种准确、鲁棒的基于DeepBoost集成学习方法的跟踪算法(DeepBoost-Tracker, DBT) .该跟踪算法采用了一种灵活、限制复杂度的目标外观模型,可实现高效处理目标复杂剧烈外观变化。该算法提出了一种联合局部全局特征的视觉表达方法以结合局部和全局视觉表达的优势。该视觉表达方法使用稀疏随机投影技术将目标主要的局部和全局空间结构信息灵活地嵌入具有多复杂度分类器成员的候选分类器集中。此外,为了避免在线学习模型过拟合,本文跟踪算法采用了一种高效的在线DeepBoost学习方法,利用其限制复杂度的特性来产生一个不易过拟合的自适应鲁棒跟踪器。实验表明,该算法的表现明显优于一般的Boosting集成跟踪算法,并取得了高精确度的跟踪结果。其次,为了在保持Boosting类型跟踪器较强自适应能力的同时增强跟踪器从跟踪失败中恢复和重启的能力,本文研究了一种歧义性正则化的多时段自适应跟踪框架,该算法框架使得跟踪系统可以从跟踪失败中恢复和重启。为了描述在跟踪过程中目标的主要外观变化,本文算法将DBT作为基础跟踪器,并使用当前学习得到的跟踪器以及从过往多时段视频帧中学习得到的跟踪器联合创建了一个跟踪器集。本文还研究了包含歧义性正则化项的选择准则损失函数,实现在每一轮跟踪中选择出具有高似然度和低歧义性的跟踪模型来确定当前帧的目标状态。当目标跟踪失败后,跟踪系统会在目标重新出现时放弃被污染的当前跟踪模型,而选择低歧义性的过往目标模型恢复和重启跟踪系统。本文研究使用高效和低计算负担的平均几何分类间隔在跟踪器自标记的数据上度量跟踪器的歧义性。实验表明,本文研究的方法可以成功地从跟踪失败中恢复和重启,并且可以处理目标剧烈复杂的外观变化,在公用测试数据集的整体和各个属性测试上都取得了排名前列跟踪结果。最后,为了提升视频监控应用中目标发现和跟踪初始化的精确度,本文还研究了具有较强通用性的基于在线子场景阴影建模和目标内边缘分析的自适应、精确运动阴影检测方法。本文分别为具有不同辐射和反射条件的子场景建立了各自的自适应在线阴影模型,以实现更加精确地描述阴影外观。为了建立子场景的在线阴影模型,本文首先计算了每个子场景运动前景区域与其被遮挡背景的在HSI色彩空间中色度、饱和度和灰度特征差累积直方图,并通过使用高斯模型来拟合这些累积直方图的最显著波峰来学习每个子区域阴影的似然模型。再者,本文研究使用目标的内边缘分析去除前景中的疑似阴影区域。最后,本文还研究使用基于局部颜色一致属性的区域生长方法来回收那些被误判为前景像素的阴影像素。实验表明,本文方法在没有光照和应用场景的先验知识的前提下,能够自适应地应对阴影外观变化和处理疑似阴影区域,其在实验测试中的表现优于其他比较的方法。更多还原

【Abstract】 As a fundamental part of computer vision, visual object tracking (VOT) is a popular studying direction in both academic and industrial community. Although significant progresses have been achieved in recent years, the tracking failure issue raised by complex and severe object appearance variations is still not solved. The object appearance variations are mainly from the situations like illumination changes, deformations, occlusions. In addition, the lack of recovery ability from the model drifting and tracking failure also degrades the tracking performance. Furthermore, the performance of moving shadow detection in the object discovering part can also affect the tracker’s initialization accuracy in vedio surveillance. The mentioned problems prevent the tracker from realizing long-term accurate persistent tracking. In this paper, we focus on alleviating these problems:We propose a novel discriminative and adaptive tracking-by-detection method based on online DeepBoost learning (DeepBoost-tracker, DBT) to improve the ability of tackling with severe object appearance variations. The proposed algorithm adopts a flexible and capacity-conscious object appearance model, which combines the strengths of both local and global visual representations. We firstly propose a joint local-global visual representation. Via applying a sparse random projection procedure on the weak classifier set, main local and global spatial structure information of the targetis flexibly embedded in the candidate classifier set with members from multiple complexity families. In addition, to avoid over-fitting our tracker adopts an effective online DeepBoost learning method (ODB). The key capacity-conscious ability of ODB can dynamically tune the complexities of selected classifiers depending on the online training set, which helps to avoid over-fitting and generate a more adaptive and robust tracker. The proposed DeepBoost-Tracker can well encode the object spatial structures and excellently handle object appearance variations. The experimental results demonstrate that our tracker outperforms the traditional boosting-style trackers and achieves very competitive tracking performance in the comparisons with the other state-of-the-art trackers.We propose an ambiguity-regularized multi-period tracking framework to enhance the recovery ability of boosting-style trackers for tracking failures, and meanwhile to keep their strong adaptive ability. For storing the object appearance changes during tracking, we build a tracker set with the current learned tracker and previous trackers learned from multiple frame periods. We incorporate an ambiguity regularization term into the loss function to select the tracker with higher likelihood and lower ambiguity. Average geometry classification margin on the tracker-labeled samples is used to measure the ambiguity of the candidate trackers. In addition, online DeepBoost algorithm is employed to strengthen the adaptive ability of base trackers. Experiments show that the proposed tracker can successfully recover from tracking failures and handle object appearance variations, and it obtains very excellent tracking results in both overall comparisons and attribute-based comparisons with the state-of-the-art trackers on the popular public test dataset.For improving the accuracy of object discovering and tracking initialization in the application of video surveillance, we propose an adaptive and accurate moving cast shadow detection method employing online sub-scene shadow modeling and object inner-edges analysis. To describe shadow appearance more accurately, the proposed method builds adaptive online shadow models for sub-scenes with different conditions of irradiance and reflectance. The models are learned by utilizing Gaussian functions to fit the most significant peaks of accumulating histograms, which are calculated from Hue, Saturation and Intensity (HSI) difference of the moving objects between the background and foreground image. Additionally, object inner-edges analysis is adopted to reject camouflages, which are foreground regions that are highly similar to shadows. Finally, the main shadow regions are expanded to recycle the misclassified shadow pixels based on local color constancy. The proposed algorithm can adaptively handle the shadow appearance changes and camouflages without prior information about illuminations and scenarios. Experimental results demonstrate that the proposed method outperforms state-of-the-art methods.更多还原