基于似物性采样和核化相关滤波器的目标跟踪算法研究

发布时间：2018-05-10 22:34

本文选题：目标跟踪 + 相关滤波　；参考：《华侨大学》2017年硕士论文

【摘要】：目标跟踪是计算机视觉领域里富有挑战性的基础研究课题,已被应用于视频监控,机器人导航和人机交互等领域。在第一帧中给定目标物体初始状态,跟踪算法的任务是在后续视频中估计目标的状态。近来,随着深度学习、相关滤波、似物性采样以及经典机器学习方法的快速发展,一系列性能优越的跟踪算法相继被提出。然而,设计一个实时的、鲁棒的和准确的跟踪器仍然是一个难题。主要是因为存在以下几个关键问题:(1)模型复杂性和跟踪实时性问题。因为跟踪对实时性要求高,现有跟踪器很难在探索更复杂的模型和保持它们的计算量尽可能小两者之间找到一个平衡点。(2)模型更新引起的漂移问题。在模型更新阶段会遇到在线训练样本不足或者自学习引入有噪声样本的情况,进而引起目标漂移。(3)遮挡或物体移出视线造成跟踪失败的问题。跟踪器可能丢失目标,通常采取滑动窗进行蛮力搜索的策略寻找目标,也不能有效地重新找回目标,不同跟踪器尝试从多方面寻找解决方案解决上述问题,其中基于核化相关滤波器的跟踪算法(Kernelized Correlation Filters,简称KCF)引起广泛关注,因为它利用循环矩阵的属性理论上分析产生许多虚拟样本,而实际上并没有增加模型复杂度,同时使用核方法和快速傅里叶变换加快运行时间。本文在KCF研究成果的基础上提出了一些有效的改进工作,主要研究内容概括如下:(1)不同于传统KCF每次仅采用单层单个核函数,本文提出一种在线协作训练的多层多核相关滤波器跟踪算法(MLMKCF)。首先,本算法将KCF单层核函数结构扩展到多层多核函数结构,利用层次结构来丰富目标表观模型。其次,提出一种有效的梯度下降法来求解核函数的权重。接着,利用两种互补的特征分别建立目标模型,两个模型在线协作更新并且使用多视角相关响应值动态融合。最后,在标注数据集(OTB-50)实验表明,本算法具有较好的准确性和鲁棒性。(2)由于目标在长时间跟踪过中存在物体遮挡、相似物体干扰和和从视野消失等问题,KCF一旦跟踪失败无法重新定位目标,本文提出一种基于似物性采样和核化相关滤波器的跟踪算法(MKCFDP)。本算法利用似物性采样方法替代传统的滑动窗检测方法,将跟踪任务被分解成跟踪和检测两个模块。跟踪模块一方面建立多层多核相关滤波器进行位置预测;另外一方面使用尺度“金字塔”估计尺度变化。检测模块是训练一个在线支持向量机(SVM)和它的历史更新存储共同组成的多专家分类器组进行重定位。当峰值旁瓣比(PSR)低于阈值时激活检测器,根据熵最小化准则选取当前最可信的分类器,对似物性采样方法Edgebox产生的粒子重新排序找出目标。将搜索区域集中在边缘突出的图像区域可以选取更有可能包含目标的候选框,优于蛮力搜索的方法。在OTB-50上的实验结果表明,本算法在效率、准确性和鲁棒性表现较好。
[Abstract]:Target tracking is a challenging basic research topic in the field of computer vision, which has been applied to video surveillance, robot navigation and human-computer interaction. Given the initial state of the target object in the first frame, the task of the tracking algorithm is to estimate the state of the target in the subsequent video. Recently, with the rapid development of deep learning, correlation filtering, physical property sampling and classical machine learning, a series of excellent tracking algorithms have been proposed. However, designing a real-time, robust and accurate tracker remains a challenge. This is mainly due to the following key problems: 1) model complexity and tracking real-time problems. Because of the high requirement of real-time tracking, it is difficult for the existing tracker to find a balance point between exploring more complex models and keeping their computational complexity as small as possible. In the updating stage of the model, the problem of insufficient training samples on line or the introduction of noisy samples from self-learning will lead to the problem of target drift. 3) occlusion or object moving out of line of sight, which results in the tracking failure. The tracker may lose the target, usually using the sliding window to carry out brute force search strategy to find the target, also can not effectively retrieve the target, different trackers try to find a solution from many aspects to solve the above problem. Kernelized Correlation filters (KCFs), a tracking algorithm based on Kernelized Correlation filters (KCFs), has attracted wide attention, because it makes use of the attributes of the cyclic matrix to generate many virtual samples in theory, but in practice it does not increase the complexity of the model. The kernel method and fast Fourier transform are used to speed up the running time. Based on the research results of KCF, some effective improvements are proposed in this paper. The main research contents are summarized as follows: (1) different from the traditional KCF, it only uses single layer single kernel function at a time. In this paper, a multi-layer multi-core correlation filter tracking algorithm based on online cooperative training is proposed. Firstly, the KCF single-layer kernel function structure is extended to the multi-layer multi-kernel function structure, and the hierarchical structure is used to enrich the target model. Secondly, an effective gradient descent method is proposed to calculate the weight of kernel function. Then, the two complementary features are used to establish the target model, the two models are updated online and the multi-view correlation response value is dynamically fused. Finally, OTB-50) experiments show that the proposed algorithm has good accuracy and robustness. Similar object interference and disappearing from the field of vision, KCF can not relocate the target once the tracking fails. In this paper, a tracking algorithm based on quasi-physical sampling and Kernel correlation filter is proposed. Instead of the traditional sliding window detection method, the tracking task is decomposed into two modules: tracking and detection. On the one hand, the tracking module establishes a multi-layer multi-core correlation filter to predict the location; on the other hand, the scale "pyramid" is used to estimate the scale change. The detection module is to train an online support vector machine (SVM) and its history update storage to make up of multiple expert classifier groups to relocate. When the peak sidelobe ratio (PSRs) is lower than the threshold, the most reliable classifier is selected according to the entropy minimization criterion, and the target is found out by reordering the particles produced by the similar physical property sampling method (Edgebox). Focusing the search area on the image region with prominent edges can select the more likely candidates to include the target, which is better than the brute force search method. Experimental results on OTB-50 show that the proposed algorithm performs well in efficiency, accuracy and robustness.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】