基于卷积神经网络的运动目标跟踪研究

发布时间：2018-08-28 08:44

【摘要】：随着信息时代的来临,运动目标跟踪成为计算机视觉领域的一个热点,在众多领域内具有广泛的应用价值。尽管已经提出众多运动目标跟踪算法,但是在实际跟踪过程中还是面临着许多困难,比如光照变化、遮挡、运动模糊、尺度变化、自身的变化等问题。因此,目标跟踪技术的发展仍然富有挑战性。深度学习理论和方法的出现为目标跟踪的研究提供了新的机遇,也是本文开展运动目标跟踪算法研究的主要理论框架。本文的主要内容如下:(1)研究了运动目标跟踪技术的基础知识。从目标跟踪的表示方法入手,了解目标跟踪的分类的基本知识和传统特征提取方法。(2)研究了卷积神经网络的基础理论。首先,在分析人工神经网络结构的基础上,介绍了卷积神经网络的结构特点、训练过程。其次,介绍了卷积神经网络特征提取的过程,对比传统特征提取和BP特征提取方法,效果明显优于这两种特征提取方法。(3)提出一种基于卷积神经网络的运动目标跟踪改进算法。基于卷积神经网络的运动目标跟踪算法是一种集深度特征提取、粒子滤波和分类器相结合的跟踪算法。首先,使用主成分分析技术(Principal Component Analysis,PCA)对局部图像数据集提取PCA特征向量,然后利用PCA特征向量初始化卷积神经网络以提取深度特征。最后,利用分类器分类、粒子滤波运动估计实现目标的识别与跟踪。实验结果表明,本文提出的运动目标跟踪改进算法可以克服跟踪过程中外界干扰和目标自身变化干扰,在准确率和成功率方面优于目前几种主流跟踪算法。
[Abstract]:With the advent of the information age, moving target tracking has become a hot spot in the field of computer vision and has wide application value in many fields. Although many moving target tracking algorithms have been proposed, there are still many difficulties in the actual tracking process, such as illumination change, occlusion, motion blur, scale change, self-change and so on. Therefore, the development of target tracking technology is still challenging. The emergence of depth learning theory and method provides a new opportunity for the research of target tracking, and is also the main theoretical framework for the research of moving target tracking algorithm in this paper. The main contents of this paper are as follows: (1) the basic knowledge of moving target tracking technology is studied. Starting with the representation method of target tracking, the basic knowledge of target tracking classification and the traditional feature extraction method are understood. (2) the basic theory of convolution neural network is studied. Firstly, based on the analysis of artificial neural network structure, the structure characteristics and training process of convolutional neural network are introduced. Secondly, the process of feature extraction based on convolution neural network is introduced. Compared with the traditional feature extraction and BP feature extraction, the effect is better than these two methods. (3) an improved algorithm of moving target tracking based on convolution neural network is proposed. The moving target tracking algorithm based on convolution neural network is a tracking algorithm which combines depth feature extraction particle filter and classifier. Firstly, the principal component analysis (Principal Component Analysis,PCA) technique is used to extract the PCA feature vector from the local image dataset, and then the convolutional neural network is initialized to extract the depth feature by using the PCA eigenvector. Finally, classifier and particle filter motion estimation are used to realize target recognition and tracking. The experimental results show that the proposed improved tracking algorithm can overcome the external interference and the change of the target itself in the tracking process, and is superior to the current mainstream tracking algorithms in terms of accuracy and success rate.
【学位授予单位】：山东科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】