结构化支持向量机目标跟踪中的特征表示与优化方法研究

发布时间：2018-03-02 07:32

本文关键词： 目标跟踪结构化支持向量机特征表示优化算法异常检测　出处：《西北农林科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：目标跟踪是计算机视觉的重要分支,融合了图像处理、模式识别、机器学习等多个学科领域的结晶,并在军事制导、视觉导航、安全监控、人机交互等方面有广阔的应用前景。近年来,结构化支持向量机(SSVM)跟踪算法由于提供了一种新颖的判别式跟踪模型及良好的性能表现而受到了广泛的关注。本文以SSVM跟踪算法为基础,从特征表示、优化方法和异常检测等三个方面进行研究,并在公开的目标跟踪基准数据库上对提出的方法进行有效性验证,主要成果如下:(1)提出了一种基于彩色Haar-like特征和选择性更新的改进Struck跟踪算法。首先使用一种彩色Haar-like特征表示方法,能够以微小的计算代价为Haar-like特征加入颜色信息。然后提出一种选择性更新模式,使得Struck跟踪器能够检测异常场景并停止对跟踪模型的更新,从而减轻模型漂移问题。在OTB50上的实验表明,使用了彩色Haar-like特征和选择性更新模式的Struck算法,在OPE、TRE、SRE三种评价方法的精准度和成功率共6个指标上,分别提升了9.1%、4.6%、7.5%、4.1%、8.3%和4.5%。(2)提出了一种对偶线性SSVM跟踪算法(DLSSVM)。本文使用了一种新的多特征目标表示方法,通过将局部秩变换(LRT)特征与Lab颜色特征相结合,能够起到突显图像细节并抑制平滑部分的作用。随后使用显式特征映射将多层特征离散化,以达到使用线性核函数近似交叉核的效果。在鲁棒性增强方面,采用多尺度目标检测算法,以适应目标的尺度变化。本部分创新性地使用了对偶坐标下降(DCD)算法求解SSVM跟踪模型,与经过显式特征映射的LRT多层特征相结合,提出的DLSSVM跟踪器性能显著优于同类型的Struck跟踪算法。(3)提出一种加权间隔SSVM跟踪模型(WMSSVM)。通过使用WMSSVM模型,能够使得跟踪器在训练更新时将样本的置信度包含在内,从而自适应地从异常场景中学习更新。在OTB100算法基准库上,WMSSVM算法在OPE和TRE的精准度和成功率等四个指标上,成绩高达82.7%、57.5%、83.5%和60.2%。综上所述,本文从三个方面对SSVM跟踪算法进行研究:在特征表示方面,使用了彩色Haar-like特征和LRT多特征;在模型求解方面,使用了DCD优化算法;在鲁棒性增强方面,使用了选择性更新、尺度估计和加权间隔模型。以上三个方面均为SSVM跟踪算法带来了不同程度的性能提升。
[Abstract]:Target tracking is an important branch of computer vision, which combines the crystallization of image processing, pattern recognition, machine learning and other disciplines, and in military guidance, visual navigation, security monitoring, In recent years, human-computer interaction and other fields have broad application prospects. Structured support Vector Machine (SVM) tracking algorithm has attracted much attention because of its novel discriminant tracking model and good performance. Based on the SSVM tracking algorithm, this paper presents the feature representation of the algorithm. The optimization method and anomaly detection are studied, and the validity of the proposed method is verified on the open target tracking benchmark database. The main results are as follows: 1) an improved Struck tracking algorithm based on color Haar-like features and selective updating is proposed. Firstly, a color Haar-like feature representation method is used. The color information can be added to the Haar-like feature at a small computational cost. Then a selective update mode is proposed, which enables the Struck tracker to detect abnormal scenes and stop updating the tracking model. Experiments on OTB50 show that the Struck algorithm based on color Haar-like feature and selective updating mode is used to evaluate the accuracy and success rate of the three evaluation methods. This paper presents a dual linear SSVM tracking algorithm called DLSSVM.A new multi-feature target representation method is used in this paper, by combining the local rank transform (LRT) feature with the Lab color feature. It can highlight the details of the image and suppress the smooth part. Then the multi-layer features are discretized by explicit feature mapping to achieve the effect of using linear kernel function to approximate crossover kernels. A multi-scale target detection algorithm is adopted to adapt to the scale change of the target. In this part, the dual coordinate descent DCD algorithm is used to solve the SSVM tracking model, which is combined with the LRT multi-layer feature which is mapped by explicit features. The performance of the proposed DLSSVM tracker is significantly better than that of the Struck tracking algorithm of the same type. (3) A weighted spaced SSVM tracking model is proposed. By using the WMSSVM model, the tracker can include the confidence of the samples in the training update. In the OTB100 algorithm benchmark database, the accuracy and success rate of OPE and TRE are as high as 82.7% and 60.2%. In this paper, SSVM tracking algorithm is studied from three aspects: color Haar-like feature and LRT multi-feature are used in feature representation, DCD optimization algorithm is used in model solving, and selective update is used in robustness enhancement. Scale estimation and weighted interval model. The above three aspects have brought about different performance improvements for SSVM tracking algorithm.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】