复杂环境下人群流量监控关键技术研究
文献类型：学位论文
| 作者 | 李静雯 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2013-06-03 |
| 授予单位 | 中国科学院大学 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 刘昌平 ; 黄磊 |
| 关键词 | 行人计数 视频监控 自适应学习 复杂环境 多摄像机 Pedestrian counting Video surveillance Adaptive learning Complex environment Multiple cameras |
| 其他题名 | Research on Technology of Pedestrian Counting in Complex Environment |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 近年来，大型公共场所的人群活动愈来愈频繁，合理管理大规模人群活动具有深远的研究意义。 智能人群监控系统能够实现全天候的人群管理，减少公共场所的安全隐患。本文旨在研究复杂环境下人群流量监控的关键技术，实现准确、鲁棒、适应性强的人群流量监控系统。本文主要工作和贡献归纳如下： 首先，针对行人流量监控领域缺乏专业数据库的现状，本文创建了CASIA行人计数数据库。该数据库包含大量实际监控场景下采集的视频、图像数据和标注信息，致力于为行人计数算法提供公开公正的评测平台。本文着重介绍了该数据库的数据采集、组成、标注、使用方法和特点。同时，归纳分析了区域行人计数和通道行人计数算法的评测指标。 基于视频的行人监控系统会面临各种不同的应用场景，视频图像中行人目标和监控背景都存在很大差异。 为了提升系统的性能，现有的行人计数系统通常对特定场景中行人和背景样本进行离线采集和模型训练，这项耗时繁琐的工作大大降低了系统的实用价值。为此，本文提出了两种基于不同模型的具备在线自适应学习功能的区域行人计数算法： 本文提出一种基于基于视觉词袋模型的在线自适应学习行人计数算法。该算法在创建初始视觉词袋模型的基础上，首先针对应用场景自动采集高置信度样本，并提出一种基于视觉词袋模型的实时更新算法，使系统能够自动获取并适应当前场景下行人目标和背景图像的特点，实现在线自适应学习功能。 本文提出一种基于混合高斯模型的在线增量学习算法。该算法通过在线采集高置信度样本实时调整模型各项参数，使模型快速适应不同应用场景，提高系统的适应能力和计数准确性，有效地避免重复离线训练过程。 针对高密度人群的监控场景，本文提出了一种能够较为准确地统计拥挤场景下区域行人数目的方法。该方法采用底层特征映射和基于目标检测的非行人前景抑制相结合的方法来实现人群块的行人数目统计，以减轻遮挡和非行人运动物体或光照变化对系统计数准确性带来的影响。 针对视野宽阔的监控场景，本文提出了一种基于多摄像机的行人计数系统。 其中，多摄像机视野范围重叠区域内的行人目标匹配是该系统需要解决的关键问题，为此本文提出一种粗匹配、细匹配两步匹配算法。 粗匹配过程得到场景间的关联信息（即图像间的映射关系）；细匹配则是抓住同一行人目标的特性进行匹配，本文提出针对行人目标的区域、轨迹、外观特征的相似性度量方法以实现不同摄像机视野重叠区域内行人目标的最优匹配，并融合多摄像机的计数结果，最终得到大场景下的行人数目。 综上所述，本文主要致力于解决复杂环境下行人流量监控的关键问题，有效提升监控系统的实用价值。 |
| 英文摘要 | In recent years, the crowd activities in large public place is more frequent. It is meaningful to do crowd management which has attracted much attention from researchers. The intelligent crowd surveillance systems can be effective to realize the crowd management 24-7, which can eliminate the drawbacks of manual monitoring and reduce the potential safety hazard of the public places. This work focuses on pedestrian counting in complex environment based on the theory of computer vision, pattern recognition and image processing. The aim of the work is to achieve an accurate, robust, adaptable people counting system. The main work and contribution of this thesis can be summarized as follows: There are still a few professional datasets for people counting. Therefore, a challenging and professional dataset is built which is named CASIA Pedestrian Counting Dataset. This dataset contains richly annotated video and images which can be used to benchmark different people counting methods.This thesis introduces the data, annotation and the usage of the dataset.Comparing with other related datasets, we discuss the characteristics of this dataset. At the same time, we also summarize the evaluation criteria for the algorithms of ROI (Region Of Interest) people counting and LOI (Line Of Interest) people counting respectively. In practice, the surveillance system is always required to handle multiple types of scenes which vary in both pedestrians and background. In order to improve the performance, most previous works used off-line training which is a time consuming task to collect enough samples in different scenes, annotate them and train classifiers or regressors. To tackle these problems, we present two kinds of adaptive self-learning people counting system which can get the number of pedestrians in a region of interest (ROI): This thesis presents a self-learning people counting system based on Bag-of-Features model. In this system, Bag-of-Features approach is used for describing the pedestrian. This thesis set some prior rules to get the positive and negative samples with high confidence and use them to do online learning. It also presents a self-learning scheme to update the model which can make the system suitable for certain particular scene. This thesis also proposes an adaptive people counting system based on Gaussian Mixture Model (GMM). Firstly, an original GMM is estimated using the Expectation maximization algorithm based on off-li... |
| 语种 | 中文 |
| 其他标识符 | 201018014628043 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6559]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 李静雯. 复杂环境下人群流量监控关键技术研究[D]. 中国科学院自动化研究所. 中国科学院大学. 2013. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。