【作者】 张元平；

【导师】 唐远炎；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2019， 博士

【摘要】 对象跟踪是一个计算器视觉领域内的基本任务,该任务旨在从图像或者视频中寻找出目标物体。该任务在自动驾驶、智能监控、人机交互以及医学成像等多种运用中具有很高的实用价值,因此大量的研究力量被投入其中。然而,在真实的对象跟踪场景中,视频目标跟踪往往会由于复杂场景条件和目标运动的随机性而变得困难。在自然条件下,目标物体经常出现光照变化、旋转、伸缩、遮挡、曝光不足、背景变化、运动的复杂性和随机性等因素,这些都会使得对象跟踪变得非常困难。虽然对象跟踪算法已经取得了丰硕的研究成果,但在理论和应用上也存在着不完善和待解决的问题。在此背景下,研究复杂场景下的对象跟踪算法在理论和应用方面都具有非常重要的意义。从神经网络的模型结构出发,对现有的模型进行优化,结合真实对象跟踪场景特性,探索新的网络模型,研究新的处理算法,提高对象跟踪的性能和效率。研究成果主要在如下几个方面体现:首先,针对复杂场景下的跟踪失效及恢复问题,提出了一种基于自组织映射网络(Self-Organizing Maps,SOM)和相关滤波器的长期记忆对象跟踪算法。在对象跟踪中,不同的视频或图像中的同一种物体常常有着不同的外观,对这些外观进行学习成为有效跟踪的关键,本算法使用具有人脑神经元的信号处理机制特点的SOM,以自适应、无监督的方式进行特征提取。同时提出了同时学习多个自适应相关滤波器来进行跟踪目标的方法,这些滤波器采样不同的更新策略,协同地进行跟踪。在大规模的基准数据集上进行的大量实验结果表明,提出的对象跟踪算法能有效地解决复杂场景中出现的严重遮挡、目标丢失、尺度变化等跟踪难题。其次,针对复杂场景下传统手工特征模型的局限导致对象跟踪的准确度问题,算法提出利用深度卷积神经网络层次特征结合相关滤波进行对象跟踪算法。在深度神经网中,前端网络层具有更多的图像纹理信息,有较高的分辨率,能够对跟踪对象进行精准的定位;而后端网络层保留了更多的图像语义信息,适合对目标形变进行检测。算法利用高层的抽象信息对目标进行粗略定位,再利用浅层的纹理细节进行精细定位,最后进行加权融合得到跟踪目标的位置。提出的自适应加权特征融合的跟踪算法结合了深度学习分层特征和相关滤波器的优势,在当前大型对象跟踪测试集上取得了很好的结果。再次,针对复杂场景下多目标跟踪的严重遮挡、尺度变化等问题,提出了一种连续两帧输入的可变形卷积层和可变形池化层的卷积神经网络用于多目标跟踪。算法在传统卷积神经网络基础上增加了可变形卷积层和可变形池化层,在卷积层和池化层中利用额外的偏移量来修改采样的位置,根据跟踪对象的实际形状进行采样。同时,算法建立了目标对象的关联轨迹,有效解决了跟踪漂移和跟踪丢失的问题。综上所述,针对对象跟踪算法的几个关键问题进行了深入的研究。对于跟踪鲁棒性及性能问题,提出利用SOM网络进行特征提取与降维,结合跟踪记忆器设计算法;对于传统手工特征模型的局限和跟踪准确度问题,提出将深度网络层特征和相关滤波器相结合的跟踪算法;此外,讨论了利用可变形卷积神经网络解决形变对象的多目标跟踪问题。提出了三种不同原创性算法,侧重于解决视觉对象跟踪算法常见的不同层面问题。大量实验结果和理论分析表明,提出的一系列视觉跟踪算法运行稳定,性能高效,显著提升了视觉对象跟踪算法的鲁棒性。更多还原

【Abstract】 Object tracking is one of the fundamental vision tasks that tries to figure out instances of several object classes from videos and images.This task has attracted a large number of attentions as it could provide the basic semantic information for numerous applications,including safety automotive systems,robotics and intelligent video surveillance.In real object tracking scenarios,the most challenges come from variations of viewpoints,illumination occlusions,non-rigid deformations and huge intra-class variations.Despite the significant progress of existing tracking methods in the last decade,object tracking remains challenging in many computer vision applications since the appearance of the object can change significantly due to pose variations,illumination changes,shape deformations,and abrupt motions.In this thesis,we address these object tracking challenges by building a hybrid neural networks based tracker to learn powerful features by taking full advantage of learning.Specifically,this thesis makes the following contributions:First,we best exploit the self-organizing map and adaptive correlation filters for for tracking failure and recovery in complex scenarios.Building upon the recent success of self-organizing maps(SOM)and correlation filter based tracking algorithm,we have extended it along several important directions.We address tracking failure dilemma by using three correlation filters that collaborate to capture the the long-term memory of the target objects appearance.In addition to the commonly used SOM features,we propose to learn correlation filters for improving targets localization accuracy.We also optimize the learning rates to improve the performance of tracking system.We explicitly handle tracking failures by incrementally learning an online detector to recover the targets.Second,based on the rich hierarchical features of the convolutional neural network,we propsose an effective visual tracking method aiming at the problem of object tracking accuracy caused by the limitation of traditional manual feature model in complex scene.Study found that the early convolutional layers include more fine-grained spatial details,which are good for precise target locating,and the last convolutional layers of convolutional neural networks remain more semantic information of target objects,which are useful for track the objects with significant appearance variations.Both layer features with semantic information and fine-grained details are exploited simultaneously for object tracking in our method.We train correlation filter on each convolutional layer and predict the target position with a coarse-to-fine searching approach.Third,we propose deformable deep convolutional neural networks to deal with the serious occlusion and scale variation of multi-target tracking in complex scenes.The new deep learning object tracking framework has innovations in multiple aspects.To make the tracking framework lighter and incorporate temporal cues,we propose a novel frame-pair based CNNs architecture.In our proposed new deep architecture,new added deformation constrained convolution and ROI-pooling layers can model the deformation of object parts with geometric constraint and penalty.With the idea of augmenting the special sampling locations in the convolution layers and pooling layers with additional offsets that are learned from the previous feature maps,both modules can easily replace their common counterparts in regular deep CNNs.Kalman Filter is used to create tracklets association that save the location and the trajectories of tracked targets,which can help us develop a more robust tracking system.These methods reduce the size of the system,improve the speed of object tracking online,and enhance the modeling ability of geometric transformation.To sum up,in this thesis,we explicitly addressed several key challenges for robust visual tracking.Combined with memory of correlation filters,SOM network is used for feature extraction and dimensionality reduction for tracking robustness and performance.Combining the deep network layer features and correlation filters,a novel tracking algorithm is proposed for the limitation of traditional manual feature model and the low tracking accuracy.In addition,the deformable convolutional neural network is introduced for multi-target tracking.Extensive experimental results on large-scale benchmark datasets demonstrate that the proposed methods in this thesis perform favorably against state-of-the-art tracking algorithms in term of effectiveness,efficiency and robustness.更多还原