多层次特征选择与特征融合在视觉跟踪中的应用

发布时间：2018-01-25 18:51

本文关键词： 计算机视觉视觉跟踪 Boosting算法 GPU加速目标特征提取　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：视觉跟踪是计算机视觉中的一个重要领域,它在视频监控、运动分析和交通监管等方面有广泛的应用。尽管目前有大量的文献给出视觉跟踪的解决方案,但由于目标姿态变换、运动模糊、遮挡以及场景中光照变化等不利因素的存在,基于视觉的目标跟踪仍然是具有挑战性的研究课题。把跟踪看成目标与背景的分类问题是解决视觉跟踪的常见方法,它不需要建立复杂模型描述目标,而是找到区分目标和背景的分类器。Grabner等人提出的基于Boosting的在线目标跟踪算法是基于分类的经典算法,该算法通过随机位置的Haar-like特征在线训练弱分类器用于选择区分效果好的特征。本文尝试使用多层次特征选择和特征融合实现目标跟踪任务,针对在线Boosting目标跟踪算法只对目标区域内位置特征作选择的问题,增加了滤波器类型的选择,提出了两层级联的Boosting改进算法;在Boosting算法框架下选择深度网络中适合跟踪的不同层次特征和不同维度特征;基于GPU的并行机制,加速两层级联的Boosting改进算法。1、本文在Boosting跟踪算法的基础上提出两层级联的Boosting跟踪方法。改进方法通过诸多滤波器模板提取目标局部特征,使用Boosting分别对目标区域内图像小块位置和它对应的滤波器类型进行选择,并且有效地融合两种特征,提升了目标跟踪的准确性。2、本文将深度神经网络中间各层的输出作为特征图谱输入Boosting算法实现目标跟踪,目的是选择适合跟踪任务的高维特征。使用Boosting分别对深度神经网络中不同层次特征和不同维度特征进行选择,并在实验结果对比中找到适合目标跟踪的特征组合方式。3、本文针对提出的两层级联的Boosting跟踪方法给出加速的方案。基于GPU的并行机制,将两层级联Boosting跟踪方法中大量繁琐的矩阵运算进行加速,提升跟踪算法的速度,增大算法的可行性。
[Abstract]:Visual tracking is an important field in computer vision. It is widely used in video surveillance, motion analysis and traffic supervision. However, due to the target attitude change, motion blur, occlusion and scene changes in the light, and other adverse factors exist. Target tracking based on vision is still a challenging research topic. It is a common method to solve the problem of target and background classification, and it does not need to establish a complex model to describe the target. The online target tracking algorithm based on Boosting proposed by Grabner et al. Is a classical algorithm based on classification. This algorithm uses the Haar-like feature of random position to train the weak classifier to select the feature with good performance. This paper attempts to use multi-level feature selection and feature fusion to achieve target tracking task. Aiming at the problem that the online Boosting target tracking algorithm only selects the location characteristics in the target region, the filter type selection is added, and a two-layer cascade Boosting improved algorithm is proposed. Under the framework of Boosting algorithm, different level and dimension features suitable for tracking in depth network are selected. Based on the parallel mechanism of GPU, the improved Boosting algorithm of two-layer cascade is accelerated. Based on the Boosting tracking algorithm, a two-layer cascaded Boosting tracking method is proposed in this paper. The improved method extracts the local features of the target by a lot of filter templates. Boosting is used to select the location of the image block and the corresponding filter type in the target region, and the two features are fused effectively, which improves the accuracy of target tracking. 2. In this paper, the output of the middle layers of the depth neural network is used as the feature map input Boosting algorithm to achieve target tracking. The purpose of this paper is to select the high-dimensional features suitable for tracking tasks. Boosting is used to select the features of different levels and different dimensions in the depth neural network. And in the comparison of experimental results to find a suitable target tracking feature combination mode. 3, this paper proposes a two-layer cascaded Boosting tracking method to accelerate the scheme. Based on the parallel mechanism of GPU. A large number of complex matrix operations in the two-layer cascade Boosting tracking method are accelerated, the speed of the tracking algorithm is improved, and the feasibility of the algorithm is increased.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】