毕业生知识库

视觉跟踪技术是计算机视觉的一个热门研究领域，在机器视觉、人机交互和视频监控等方面有着广泛应用。近年来在国内外研究者的努力下，各类视觉跟踪算法层出不穷，在速度和精度上已经取得了相当大的进步。相关滤波跟踪算法基于岭回归框架，利用循环采样和快速傅里叶变换实现封闭解的密集匹配跟踪，在速度和鲁棒性上均具有相对出色的性能，逐渐成为了视觉跟踪领域的主流算法。为进一步提升相关滤波跟踪算法的性能，各种正则化思想被相继引入，但破坏了相关跟踪算法的封闭解，降低了实时性，阻碍了其在无人机、移动机器人等低算力约束场景下的应用潜力。本文主要基于相关滤波算法，研究低算力应用场景下的视觉目标跟踪技术，在提高视觉跟踪算法的性能的同时保持实时性，达到性能和效率之间的平衡。论文的主要工作包含以下四个部分：
（1）相关滤波与尺度自适应技术研究。针对相关滤波跟踪算法无法自适应目标尺度变化的问题，总结了现有相关跟踪中各类尺度自适应技术，分析了各种方案的性能、效率及其优缺点，通过在公开数据集 OTB-2015 上的对比实验，确定尺度自适应方案。
（2）高效的跟踪特征表达与聚合方法研究。判别式目标跟踪需要对目标和背景提取高效且鲁棒的特征。针对低算力约束场景下的视觉跟踪应用需求，借鉴mean-shift 跟踪算法中权值图的思想，提出了 1 通道的颜色比率（CR）特征以替换 10 通道的颜色命名（CN）特征，并采用 13 通道的 HOG 特征替换常用的 31通道特征，以进一步提升算法的实时性。另外，在特征组合中对不同类型特征乘以不同权重来控制其在最终特征表达中的比重，以更好地发挥各类特征的判别力，使得跟踪器的性能得到了进一步提升。实验结果表明，所提算法（简称为CRCF）同时实现了性能和效率的显著提升。
（3）模型更新机制研究。跟踪算法通常需要在跟踪过程中对模型进行更新，以保留目标历史信息的同时适应目标的形状变化。现有大部分基于相关滤波的跟踪算法采用滑动平均更新的方式来更新模型，但这种方式在目标被遮挡时容易造成模型污染而导致跟踪漂移，在目标发生快速形变时又易由于更新不够而导致跟踪失败。结合相关滤波领域的最新进展，提出基于差异哈希距离度量的在线高斯混合聚类的模型更新机制，在线维护目标代表性历史形态的样本库，并通过样本库训练判别能力和泛化能力更强的相关滤波器。实验结果表明，所提算法（简称为 CRCF_ATU ）在不降低效率的情况下相比 CRCF 取得了进一步性能提升。
（4）长时跟踪算法研究。长时跟踪要求跟踪器准确检测跟踪失败和目标消失等情况，并在目标再次出现时重新定位目标。结合上文所述模型更新机制，提出了一种更准确的跟踪状态判别机制。在判断跟踪失败时，结合样本库训练背景感知相关滤波器，实现对目标的全局重检测。实验结果表明，所提长时跟踪算法（简称为 CRCF_LCT ）可以准确判断跟踪状态，并在目标重新出现后找回目标，大幅提高算法在长时视频序列上的表现。
CRCF、 CRCF_ATU、 CRCF_LCT 分别适用于超短时跟踪、短中时跟踪、长时跟踪应用场合，可以为检测跟踪融合、单目标可靠跟踪、主动视觉跟踪等不同场景下的应用提供适配的跟踪模块，具有重要的应用前景。

Visual tracking has been a popular problem in computer vision with widespread applications in robotics, human-computer interaction and video surveillance. In recent years, various tracking algorithms have been developed and considerable progress has been made in terms of speed and accuracy thanks to the effort of researchers at home and abroad. Correlation filters perform dense-sampling based on ridge regression with Fast Fourier Transform by exploiting circulant structure. Correlation filters based trackers perform favourably in terms of speed and robustness and have become mainstream branch in the field of visual tracking. To further enhance the performance of correlation filter based trackers, various types of regularization are introduced. However, the introduction of regularization breaks the closed-form solution of correlation filters and reduce the speed, which restricts applications in some computation restricted scenarios such as UAV and mobile robots. This thesis focuses on visual tracking based on correlation filters in computation restricted applications to improve tracker performance with real-time speed and achieve balance between performance and efficiency. Our work mainly includes the following four parts:
(1)Research on correlation filter and scale adaptive methods. Considering the problem of scale variation in correlation filter tracking, we review current scale adaptive methods in correlation filter tracking and analyse their advantages and disadvantages in terms of performance and efficiency. By conducting comparative experiments on OTB-2015, we determine the scale adaptive method of this thesis.
(2)Research on efficient feature representation and aggregation. Discriminative tracking requires to extract efficient and robust feature representation from target and background. Considering the visual tracking demand in computation restricted applications, we borrow the essential idea of weight image in mean-shift based trackers and propose 1-channel color ratio(CR) feature to replace 10-channel color-naming(CN) feature and use 13-channel HOG feature to replace common 31-channel HOG feature to improve efficiency. In addition, during feature concatenation process, we weight between different types of features to control their importance in final feature representation in order to better exert the discriminative power of the features and further improve the tracker performance. Experimental results demonstrate that the proposed method(denoted as CRCF) brings notable improvement in performance and efficiency.
(3)Research on model update scheme. Trackers require to update the model in tracking process to adapt to target appearance variation and reserve historical target appearance. Most correlation filter based trackers update the model with moving average. However, this update scheme usually leads to tracker drift due to model contamination under occlusion and is prone to fail under large target appearance variation due to limited update. Considering latest progress in correlation filter tracking, we propose an update scheme with online Gaussian mixture model and difference hashing distance measurement to maintain a training set of representative target historical appearance. Correlation filter with better discriminative power and generalization ability is trained on the training set. Experimental results demonstrate that the proposed method(denoted as CRCF_ATU) brings performance gain compared with CRCF without reducing efficiency.
(4)Research on long-term tracking. In long-term tracking, the tracker is required to accurately detect tracking failure and target disappearance. When the target reappears, the tracker is required to localize the target. With aforementioned model update scheme, we propose a more accurate tracking status judgement scheme. When tracking failure is detected, background-aware correlation filter(BACF) is trained on the training set to perform global re-detection. Experimental results demonstrate that the proposed method(denoted as CRCF_LCT) can accurately re-detect target after target disappearance and brings notable performance gain in long-term sequences.
CRCF, CRCF_ATU and CRCF_LCT respectively apply to short-term tracking, middle-term tracking and long-term tracking, providing adaptable visual tracking modules for various applications such as detection and tracking fusion, reliable single target tracking and active visual tracking with considerable prospect.