基于局部特征提取的目标检测与跟踪技术研究

发布时间：2018-03-10 05:04

本文选题：局部特征提取　切入点：支持向量机　出处：《中国科学院大学(中国科学院光电技术研究所)》2017年硕士论文　论文类型：学位论文

【摘要】：目标检测与跟踪作为计算机视觉研究领域的一个重要组成部分,未来有望广泛应用于运动的识别、自动监督、视频检索、人机交互等诸多现代技术领域。而传统的目标检测与跟踪算法一般只采用图像底层的视觉特征来构建目标描述子,获得目标外观的描述,来进行目标检测与跟踪;或者通过简单的背景建模方式来区分前景和背景,从而获取目标的位置信息,实现目标的检测与跟踪。然而在实际过程中,除了存在目标自身姿态、尺度、旋转、平移变化,还存在光照变化、复杂背景、目标遮挡等挑战。由于底层特征对这些变化不具有不变性,无法精确的描述目标,造成跟踪器失效。因为局部特征对目标检测与跟踪过程中的变换具有良好的不变性,所以通过提取目标的局部特征,构建鲁棒的目标描述子并结合鲁棒的目标检测与跟踪框架成为了目标检测与跟踪技术研究中的一个热点。在目标检测方面,首先分析了HOG-PCA描述子和支持向量机目标检测方法的优势与不足。针对在提取HOG局部特征时采用传统四象限组合成块方式,构建的目标描述子不能精确的描述目标,并对后续提取特征主元没有优势的不足,提出了一种改进HOG-PCA特征描述子的方法,在提取HOG特征时采用极坐标组合块方法代替传统四象限的方法,构建能更精确表示目标的描述子。另一方面通过循环采样的方式代替传统的随机采样方式,构建用于SVM训练的正负样本,使训练器训练的结果更加精确。通过提取正负样本改进后的HOG-PCA描述子,进行主成分分析,然后作为样本数据,用于支持向量机分类器的训练,最后将训练好的分类器用于目标的检测。在目标跟踪方面,同样从提取目标的局部特征,构建鲁棒的目标描述子出发,针对传统的空时上下文跟踪器中只使用高斯加权的底层灰度特征来获得目标的特征描述,在复杂场景下,不能获到鲁棒跟踪结果的不足。本文通过提取目标图像的局部圆域混合块LBP特征,构造图像的响应图,获得了目标的外观描述,然后在贝叶斯框架下对目标和它的局部上下文区域的时空关系进行建模,提出了一种利用目标周围上下文视觉显著信息进行跟踪的算法。本文在目标检测方面使用了INRIA行人数据集来对改进后HOG-PCA特征的SVM目标检测方法进行了全面的测试,将其与原来的算法进行了比对,实验结果表明改进后的算法能够明显的降低目标的误检率。在目标跟踪方面使用了大量的经典跟踪视频集对算法进行了全面的测试,也加入了另外三种经典跟踪算法的结果来进行比较与分析,实验表明本文算法的性能要胜于其他跟踪算法。最后,对本文的研究内容进行了全面的总结,并为后续的研究工作提供了思路。
[Abstract]:As an important part of computer vision research, target detection and tracking is expected to be widely used in motion recognition, automatic monitoring and video retrieval in the future. The traditional target detection and tracking algorithms only use the visual features of the bottom layer of the image to construct the object descriptor and obtain the description of the appearance of the target to detect and track the target. Or it can distinguish the foreground and background by simple background modeling method, so as to obtain the position information of the target, and realize the detection and tracking of the target. However, in the actual process, there are not only the changes of the target's own attitude, scale, rotation and translation, but also the change of the target's attitude, scale, rotation and translation. There are also challenges such as illumination changes, complex backgrounds, object occlusion, and so on. Because the underlying features are not invariant to these changes, it is impossible to accurately describe the target. Because the local feature has good invariance to the transformation in the process of target detection and tracking, so by extracting the local feature of the target, Constructing robust target descriptor and combining robust target detection and tracking framework has become a hot topic in the research of target detection and tracking technology. Firstly, the advantages and disadvantages of HOG-PCA descriptor and support vector machine (SVM) are analyzed. In order to extract the local features of HOG, the target descriptor can not accurately describe the target by using the traditional four-quadrant combination method. In this paper, an improved method of HOG-PCA feature descriptor is proposed, in which polar coordinate combination block method is used to replace the traditional four-quadrant method in HOG feature extraction. On the other hand, instead of traditional random sampling, positive and negative samples for SVM training are constructed. By extracting the improved HOG-PCA descriptor of positive and negative samples, the principal component analysis (PCA) is used as sample data to train the classifier of support vector machine (SVM). Finally, the trained classifier is used for target detection. In the aspect of target tracking, a robust target descriptor is constructed by extracting the local features of the target. In the traditional space-time context tracker, only Gao Si's weighted low-level gray features are used to obtain the feature description of the target. The result of robust tracking can not be obtained. In this paper, by extracting the local circular region mixed block LBP features of the target image, the response graph of the image is constructed, and the appearance description of the target is obtained. Then the spatiotemporal relationship between the target and its local context is modeled under the Bayesian framework. In this paper, we propose a tracking algorithm using visual salient information about the context around the target. In this paper, we use INRIA pedestrian data set to test the improved SVM target detection method based on HOG-PCA features. Compared with the original algorithm, the experimental results show that the improved algorithm can significantly reduce the false detection rate of the target. In the aspect of target tracking, a large number of classical tracking video sets are used to test the algorithm. The results of the other three classical tracking algorithms are compared and analyzed. The experimental results show that the performance of this algorithm is better than that of other tracking algorithms. Finally, the research content of this paper is summarized comprehensively. It also provides the train of thought for the following research work.
【学位授予单位】：中国科学院大学(中国科学院光电技术研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】