面向真实场景的大范围视觉群体目标分析

CASIA OpenIR > 毕业生 > 博士学位论文

	面向真实场景的大范围视觉群体目标分析
	曹黎俊
	2016-05-30
学位类型	工学博士
中文摘要	智能视频监控作为公共安全管理的一个有效手段，越来越受到各方的重视。随着国内各地平安城市、天网工程等一大批的安防摄像机的建设，大范围覆盖、多摄像机协同场景下的视觉群体目标的分析技术，成了智能视频监控的一个重要研究方向。群体目标分析的内容包括目标的轨迹、密度、数量和流量等，在真实视频监控应用中，常常因为诸如复杂的现场环境、恶劣的天气、变化的光照、各种噪声的干扰等因素，难以推广使用。而在多摄像机协同分析大范围场景下的群体目标时，不但对算法的场景适应性要求更高，协同分析目标轨迹、密度和流量等，也变得更加困难。本文以实际应用为背景，以群体目标在多摄像机场景下的轨迹跟踪和密度、流量分析为切入点，深入研究群体目标在真实大范围场景下的分析方法，形成完整的大范围监控场景下的群体目标分析方法流程，以及相关的应用系统。论文的主要工作和贡献如下： 1. 提出了一种基于均衡全局图模型的多摄像机目标轨迹跟踪方法，克服了传统两步式多摄像机目标跟踪方法的缺陷（要求首先获得单摄像机目标轨迹或者目标检测结果），使得其实用性大大提高。该方法通过提取目标的稳定轨迹片段，使用全局图模型把单摄像机轨迹跟踪和跨摄像机轨迹关联融合在一个框架下进行优化，并且对不同摄像机下的目标表观和运动特征的相似度进行均衡化处理，最终明显提升了跨摄像机轨迹关联的准确性。通过对单摄像机和跨摄像机下轨迹跟踪结果的综合评价，表明该方法在大范围轨迹分析时，特别是单摄像机目标跟踪性能有限时，有较好的性能。 2. 提出了一种场景通用性较好的群体密度估计方法，解决了传统方法场景适应性差、参数调节复杂的问题，并且实现了在大范围场景下，通过多个摄像机对局部人群密度的分析和关键点位人群密度的预测。该方法使用马赛克图像累积差分特征提取群体聚集区域，并对其进行透视变换矫正用以估计人群密度。通过对人群数量和运动速度的估计，可以实现预测指定点位的人群密度。经过与现有的方法对比，以及在真实场景中的应用后发现，该方法无论在不同的场景中，还是在不同的人群密度下，不需要调节过多参数即可保持较为有效和稳定的效果。 3. 提出了一种基于深度神经网络模型的群体目标流量分析方法，并创建了十万级的相关数据库用于训练，相比于传统方法，提升了对新场景的适应性和密集人群流量统计的稳定性。该方法借鉴了图像序列采样的思想，提取了视频的原始采样图和光流特征采样图，并使用多个深度神经网络模型对采样图进行特征提取，以此来估计群体流量。创建的群体目标流量估计数据库，不但包含实际场景中的不同角度和不同规模的群体流量，并且涵盖了不同的天气情况和时间段。实验表明，传统的方法只能在特定的场景，流量较少时发挥作用，而本文提出的方法可以较好的估计嘈杂室外环境中的大规模群体目标流量。
英文摘要	As one of the effective ways of public security management, intelligent video surveillance has been attracting attentions from different communities. With the construction of a series of public security related projects, such as Safe City and Skynet project, group visual target analysis, especially with large scale coverage and multi-field collaborative scenario, has become an important research area. Group target analysis involves the target trajectory, density, number and flow. In real world applications of video surveillance, group target analysis is difficult to employ because of complicated background, changeable weathers, illumination variations and noise clutters. The group target analysis for large area scenes with multi-cameras raises higher demands for the adaptability of algorithms, and faces more difficulties in trajectory, density and flow analysis. Based on real world applications, and taking the target tracking, density and flow analysis as the research points, this thesis studies the methods of group target analysis for large scale scenes. Based on the above works, a systematic procedure of group target analysis and an application system are further constructed. The main content and contributions of this thesis are listed as follows. 1. The thesis presents an equalized graph model based multi-camera target trajectory analysis method, which overcomes the weakness of traditional two-step trajectory analysis (which requires the tracking results of single camera beforehand), and improves the availability significantly. By extracting tracklets of target, the method associates the tracklets in a global graph model and performs equalization on the appearance and motion feature similarities, which significantly improves the association accuracy. The comprehensive evaluation of both single and multi camera scenarios demonstrate that the proposed method performs excellent in large area trajectory analysis, especially when the ability of single camera tracking algorithm is limited. 2. The thesis presents a group density estimation method with better generalization ability, and overcomes the problems of new scene adaptation and complex parameters adjusting, and implements an algorithm for local group density analysis in multi-camera settings under large area scenes. The method employs the accumulated mosaic image difference as the feature representation for crowded area extracting, and applies the perspective transformation for correcting the images. By estimating the number and velocity of group targets, the method can predict group density in specific areas. Comparison results with the existing methods in real scenes demonstrate the stability of our method without much adjusting for parameters. 3. The thesis presents a deep neural network model based group flow analysis, and constructs a data set containing more than 100 thousand images for training. Compared with traditional methods, it improves the stability and adaptability for new and crowded scenes. The method is motivated by the sequential sampling of images, which extracts the original and optical flow sampling images, and employs multiple deep neural networks for feature extraction to estimate the group density. The constructed data set not only contains groups captured from different angle and has different scales, but also covers different weather and time. Experimental results demonstrate that our method works better in crowded outside scenes and in larger areas, while traditional methods can only work under specific scenes, when the group is not very crowded.
关键词	大范围场景群体目标轨迹关联密度估计流量分析
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/11815
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	曹黎俊. 面向真实场景的大范围视觉群体目标分析[D]. 北京. 中国科学院大学,2016.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
面向真实场景的大范围视觉群体目标分析 .（32706KB）	学位论文		限制开放	CC BY-NC-SA