基于自组织神经网络的目标轨迹分析研究

发布时间：2018-01-13 00:07

本文关键词：基于自组织神经网络的目标轨迹分析研究　出处：《昆明理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着社会的发展,以视频图像为数据基础的自动化视频轨迹分析在各领域具有重要的研究意义,包括视频图像有效信息的提取,数据信息的学习,建立合适的数据模型进行检查等。视频数据是物体运动轨迹的数据基础,使用数字图像处理的相关方法,过滤掉与提取运动轨迹无关的背景图像,提取出研究需要的相关物体的运动轨迹,是进行运动物体轨迹分析的基础。本文应用有效的方法提取轨迹数据的主要特征,对轨迹进行聚类研究,通过聚类结果分析轨迹的运动特征。本文首先论文介绍了运动目标轨迹的研究背景和研究现状,然后介绍了从视频中提取运动目标轨迹的常用方法,分析了各个提取方法的优点和不足。接下来对于轨迹数据集中的短轨迹,应用改进的灰色马尔科夫预测法,补全了轨迹点,预测了短轨迹未来时间段的运动趋势。接下来介绍了流形学习的常用方法,分析了不同方法的适用数据集的特征。针对经典的欧式距离度量缺乏对轨迹特征的描述,不能准确描述二维轨迹间流形特征的差异,本文将轨迹数据映射到向量场中,建立轨迹数据的向量场,保留了主要特征信息,实现了轨迹数据的可视化表示,将轨迹数据的形状特征直观的表示出来。最后,介绍了自组织神经网络和kmeans聚类。对大西洋飓风轨迹数据作为实验数据,对聚类结果进行分析,验证了本文算法的有效性。
[Abstract]:With the development of society, automated video trajectory analysis based on video image has important research significance in various fields, including the extraction of effective information of video image and the learning of data information. The video data is the data base of the moving track of the object. Using the related method of digital image processing, the background image which is independent of extracting the motion track is filtered out. Extraction of the trajectory of related objects is the basis of trajectory analysis. This paper applies an effective method to extract the main features of trajectory data and cluster the trajectory. Firstly, this paper introduces the research background and research status of moving target trajectory, and then introduces the common methods of extracting moving target trajectory from video. The advantages and disadvantages of each extraction method are analyzed. Then for the short trajectory in the track data set, the improved grey Markov prediction method is applied to complement the trajectory points. The movement trend of the short trajectory in the future is predicted. Then the common methods of manifold learning are introduced. The characteristics of the suitable data sets of different methods are analyzed. Because the classical Euclidean distance measurement lacks the description of the trajectory features, it can not accurately describe the difference of the manifold characteristics between the two dimensional trajectories. In this paper, the trajectory data is mapped to the vector field, the vector field of the trajectory data is established, the main characteristic information is retained, and the visual representation of the trajectory data is realized. Finally, self-organizing neural network and kmeans clustering are introduced. The Atlantic hurricane track data as experimental data, clustering results are analyzed. The validity of the proposed algorithm is verified.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】