SIA OpenIR  > 光电信息技术研究室
面向复杂任务的目标跟踪技术
Alternative TitleResearch on target tracking technology under complex task
陈法领
Department光电信息技术研究室
Thesis Advisor丁庆海
Keyword计算机视觉 目标跟踪 时空上下文 相关滤波 卷积神经网络
Pages122页
Degree Discipline模式识别与智能系统
Degree Name博士
2020-11-26
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract目标跟踪问题是当前计算机视觉研究领域的一项重要课题,它依据目标在视频序列起始帧的先验信息,提取目标在视频序列后续帧的特征表示,利用目标在视频序列中的时间和空间关联性,估计出目标的位置、尺度、速度和轨迹等状态信息,可以为更高层次的计算机视觉任务提供重要的信息输入。目标跟踪在智能监控、人机交互、智能交通、机器人导航、辅助医学诊断和精确制导等诸多领域中都具有广泛的应用,因此,该项研究工作具有十分重要的理论意义和实际的应用价值。在国内外研究人员数十年的不懈努力下,目标跟踪技术已经取得了很大进展,但在复杂任务中,由于目标自身及其所处环境的非受控特点,容易受到目标形状变化、尺度变化、被遮挡以及光照变化和背景干扰等多种不利因素的影响,实现稳定可靠的目标跟踪仍面临着严峻的挑战。本文主要面向复杂任务中的目标跟踪问题,分别针对不同的应用场合开展了目标跟踪算法的研究和探索,同时为算法的工程应用进行了目标跟踪系统的设计与实现。本文的主要研究内容和取得的成果如下:1.针对算力有限的低成本应用场合,提出了一种基于时空上下文的抗遮挡目标跟踪算法,解决了时空上下文目标跟踪模型的跟踪点漂移问题。首先,通过提取目标及其周围上下文区域的颜色名(Color Name, CN)特征并进行自适应降维,在提高算法跟踪性能的同时降低了对计算资源的要求。其次,根据跟踪过程中目标模板的相关系数变化判断目标外观是否发生变化,在目标外观因遮挡等干扰发生轻微改变时,通过Kalman滤波对目标位置进行修正,并降低跟踪模型的更新速度;而当目标外观发生剧烈变化时,则利用Kalman滤波得到的目标状态预测目标位置,并停止更新跟踪模型,在目标外观恢复后,重新捕获目标并进行跟踪。实验结果表明,所提算法在出现了目标形变和背景干扰等不利因素时,仍具有较好的跟踪稳定性,并具有较强的抗遮挡能力。2.针对高性能的目标跟踪应用场合,提出了一种自适应特征融合的多尺度目标跟踪算法,解决了相关滤波目标跟踪模型的特征融合和尺度估计问题。首先,分别提取目标候选区域的梯度方向直方图(Histogram of Oriented Gradient, HOG)特征和CN特征进行目标表征,根据所提出的权重自适应融合策略,将这两种特征的相关滤波响应进行加权融合,完成目标的位置估计。然后,以此为中心进行多尺度采样,利用尺度相关滤波以及相邻帧之间目标尺度变化的先验分布,实现了对目标尺度变化的最大后验概率估计。最后,根据融合后响应图的可信度完成对跟踪模型的更新。实验结果表明,所提算法不仅能够较好地适应目标的尺度变化,而且相对于其它对比算法也具有较高的跟踪精度和成功率。3.着眼于未来的技术储备,提出了一种基于多层卷积特征决策融合的背景感知目标跟踪算法,充分发挥了卷积神经网络模型提取的卷积特征的目标表征能力。首先,对不同层级卷积特征的目标表征特点进行了分析,然后在背景感知滤波模型框架下选择多个层级的卷积特征分别训练得到多个弱跟踪器。为了适应目标跟踪过程中的变化,根据各个弱跟踪器的决策损失变化,自适应地调整它们的决策权重,融合得到一个强跟踪器完成对目标的位置估计。与此同时,为了进一步提升算法的跟踪性能,提出了一种目标重检测机制,通过在“后台”运行一个目标跟踪状态评估模块,分别对跟踪和重检测的结果进行可靠性判别,以此完成对目标的稳定跟踪和可靠更新。实验结果表明,所提算法不仅在目标发生了旋转和形状变化时仍具有较高的跟踪性能,而且能够适应目标被遮挡以及光照变化和背景干扰等多种挑战因素。 4.针对目标跟踪算法的工程应用问题,设计并实现了一种面向目标跟踪的高速并行信息处理系统。在分析了现阶段实现目标跟踪算法的各种处理器特点的基础上,以高性能多核DSP和超大规模FPGA为核心实现了一种具有并行处理体系结构的嵌入式实时目标跟踪系统。在DSP和FPGA之间设计了高速串行数据通道,实现了目标跟踪算法运行过程中大容量图像数据的低延时传输。完成了图像预处理、特征提取和快速傅里叶变换等关键模块的硬件加速引擎设计,为目标跟踪算法的工程应用奠定了坚实的基础。
Other AbstractTarget tracking is an important research issue in computer vision currently. Based on the prior information of the target in the start frame, the features of the target are extracted in subsequent frames, and the temporal and spatial relationship of the target in the video sequence are combined to estimate its position, scale, speed, trajectory and other status information, meanwhile these important information will be provided for high-level computer vision tasks. Therefore, target tracking has been used widely in various applications such as intelligent surveillance, human-computer interaction, intelligent transportation, robot navigation, auxiliary medical diagnosis and precision guidance, so the research on target tracking has important theoretical significance and application value. Although it has been made great progress by the efforts of many domestic and foreign researchers in the fewer decades, however, in complex task, due to the target and the environment are uncontrolled, target tracking is susceptible to unfavorable factors such as target deformation, scale variation, occlusion, and illumination variation and background clutter, so it is still a very challenging task to achieve stable and reliable target tracking. The dissertation is mainly oriented for the target tracking problem under complex task. For different applications, the target tracking algorithms are researched and developed separately. Meanwhile, a target tracking system is designed and implemented for their engineering applications. The main contents and contributions of this dissertation are as follows. For a low-cost application with limited computing resource, an anti-occlusion target tracking algorithm based on spatio-temporal context is proposed to tackle the tracking drift of the spatio-temporal context target tracking model. Firstly, the color name features are extracted of the target and its surrounding context and their dimensions are reduced adaptively for target tracking, which improves the tracking performance and reduces the requirement of computing resource. Secondly, according to the change of target templates’ correlation coefficient in tracking process, the appearance variation of the target is evaluated. When the target appearance is slightly changed due to occluded, the Kalman filter is used to correct the target position and the update speed of the tracking model is reduced. When the target appearance is severely changed, the Kalman filter is used to predict the target position with its previous state and the update of the tracking model is stopped. When the target appearance recovers, it will be recaptured and tracked again. The experimental results demonstrate that the proposed algorithm has better tracking stability under the influences of target deformation and background clutter, and it has strong anti-occlusion ability also. Aiming at a high-performance target tracking application, a target tracking algorithm with adaptive features fusion and multi-scale estimation is proposed to solve the problem of features fusion and scale estimation of the correlation filter target tracking model. In the first place, the histogram of oriented gradient features and the CN features are extracted at the target candidate area for target characterization, and according to the proposed weighted adaptive fusion strategy, the two correlation filter responses are combined to complete the target position estimation. In the second place, multi-scale image patches are sampled at the center of the evaluated position, then combining a scale correlation filter and the prior distribution of target scale variation between two adjacent frames, the target scale variation is estimated with the maximum posterior probability. Finally, the tracking models are updated according to the confidence of the fused response. The experimental results demonstrate that the proposed algorithm can not only adapt to the scale variation well, but also has higher tracking accuracy and success ratio compared with other comparison algorithms. With a view to future technology, a background-aware target tracking algorithm based on multi-layer convolutional features decision fusion is proposed to make full use of the target representation ability of convolution features, which are extracted by the convolutional neural network model. Firstly, the target representation ability of different levels of convolutional features is analyzed, and the multi-layer convolutional features are employed to construct several weak trackers under the background-aware filter framework. In order to adapt to the changes in target tracking process, their decision weights are adjusted adaptively according to the decision loss of each weak tracker, and a strong tracker is merged in this way to estimate the target position. At the same time, in order to improve the tracking performance of the algorithm further, a target re-detection mechanism is proposed. By running a target tracking status evaluation module "background", the reliabilities of the tracking result and the re-detection result are assessed, then the stable tracking and the reliable update are completed in this way. The experimental results demonstrate that the proposed algorithm not only has high tracking accuracy when the target is rotation and deformation, but also can overcome many challenging factors such as occlusion, illumination variation, and background clutter under the complex task. For the purpose of the engineering application of the target tracking algorithm, a high-speed parallel information processing system is designed and implemented. After analyzing the characteristics of various processors which can be used for realizing the target tracking algorithm currently, taking high-performance multi-core DSP and ultra-large-scale FPGA as the core, an embedded real-time target tracking system with parallel process architecture is implemented. Several high speed serial data transmission channels are designed between the DSP and FPGA in this system, so that the low-delay transmission of large capacity image data is complemented during the process of target tracking. The hardware acceleration engines of essential modules such as image preprocessing, feature extraction and fast Fourier transform has been accomplished, which build a solid foundation for the engineering application of the target tracking algorithm.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/27981
Collection光电信息技术研究室
Recommended Citation
GB/T 7714
陈法领. 面向复杂任务的目标跟踪技术[D]. 沈阳. 中国科学院沈阳自动化研究所,2020.
Files in This Item:
File Name/Size DocType Version Access License
面向复杂任务的目标跟踪技术.pdf(6052KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[陈法领]'s Articles
Baidu academic
Similar articles in Baidu academic
[陈法领]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[陈法领]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.