基于特征加权的快速压缩感知目标跟踪算法研究

发布时间：2018-10-16 12:40

【摘要】：目标跟踪具备重要的科研价值,且被普遍应用于视频监控、智能交通等之中,近年来一直都是机器视觉研究当中的热点问题。针对目标跟踪问题,当下已提出众多研究算法。虽然这些跟踪算法具有一定的鲁棒性,但由于被跟踪目标外观在跟踪过程中易因受到光照、遮挡等外界因素的影响发生变化,使得视觉目标跟踪仍然是一个难点问题。压缩感知理论作为信号领域的研究热点,因其具有简单、高效、实时等优点,在跟踪领域也获得了广泛使用。由研究证明,通过压缩感知技术获取的低维压缩信号可基本保留原信号的全部特性,从而可在保证计算精度的前提下大幅提高运算效率。利用压缩感知理论设计的跟踪算法能够获取快速、鲁棒的跟踪结果。但此类算法在获取图片的多尺度表示,生成目标高维特征时,未考虑随着滤波器尺寸的增大,由滤波产生的有效特征逐渐减少。同时算法在度量目标与候选样本之间的相似性时,仅将每个弱分类器分类结果进行简单叠加,处理策略过于单一。当目标因受到光照、遮挡等外界因素的影响外观发生变化时,易降低跟踪精度使得结果出现偏差。本文针对上述问题,提出特征加权的快速压缩感知目标跟踪算法,通过生成准确的高维特征及更加高效的相似性度量机制来获取跟踪,具体工作如下:(1)研究学习压缩感知理论,深入了解信号稀疏表示理论、设计观测矩阵的方法和其常用分类、重构信号算法以及压缩感知降维等。实验论证压缩跟踪算法的有效性,发现此类压缩跟踪算法中存在的不足并进行分析与解决;(2)依据滤波器尺寸提出自适应加权策略来提取样本的多尺度特征,进而生成高维特征描述。该方法能够解决随着滤波器尺寸的增大,导致获取到的样本有效特征逐渐减少的问题,确保跟踪使用特征的准确性;(3)兼顾考虑每个弱分类器的叠加和与候选样本各维压缩特征被分类为目标压缩特征的可能性,提出了高效的相似性度量机制。该方法解决了之前衡量样本与目标相似性方式单一的问题。本文与当下现有的此类压缩跟踪算法在多个测试视频序列中进行对比,结果可知,本文所提算法具有更优的跟踪性能。
[Abstract]:Target tracking has important scientific research value, and is widely used in video surveillance, intelligent transportation and so on. In recent years, it has been a hot issue in machine vision research. Many algorithms have been proposed for target tracking. Although these tracking algorithms have some robustness, the visual target tracking is still a difficult problem because the appearance of the target is easily affected by illumination, occlusion and other external factors in the tracking process. As a research hotspot in signal field, compression sensing theory has been widely used in the field of tracking because of its advantages of simplicity, efficiency and real-time. It is proved by the research that the low dimensional compression signal obtained by compression sensing technology can basically retain all the characteristics of the original signal, thus greatly improving the computational efficiency on the premise of guaranteeing the calculation accuracy. The tracking algorithm based on compressed sensing theory can obtain fast and robust tracking results. However, in obtaining the multi-scale representation of the image and generating the high-dimensional feature of the target, this algorithm does not consider that the effective feature generated by the filter decreases gradually with the increase of the size of the filter. At the same time, when measuring the similarity between the target and the candidate samples, the algorithm simply superposes the classification results of each weak classifier, and the processing strategy is too single. When the external factors such as illumination and occlusion affect the appearance of the target, it is easy to reduce the tracking accuracy and make the results deviate. In order to solve the above problems, this paper proposes a feature weighted fast compressed perceptual target tracking algorithm, which obtains tracking by generating accurate high-dimensional features and more efficient similarity measurement mechanism. The main work is as follows: (1) Learning compressed perception theory is studied. We deeply understand the theory of sparse representation of signals, design the method and classification of observation matrix, reconstruct the signal algorithm and reduce the dimension of compression perception. Experiments demonstrate the effectiveness of the compression tracking algorithm, find the shortcomings of the compression tracking algorithm and analyze and solve it. (2) according to the filter size, an adaptive weighting strategy is proposed to extract the multi-scale feature of the sample. Then high dimensional feature description is generated. This method can solve the problem that the valid features of the obtained samples decrease gradually with the increase of filter size. (3) considering the superposition of each weak classifier and the possibility that each dimension of the candidate sample is classified as the target compressed feature, an efficient similarity measurement mechanism is proposed. This method solves the problem that the similarity between sample and target is simple. Compared with the existing compression tracking algorithms in several test video sequences, the results show that the proposed algorithm has better tracking performance.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】