高密度动态人群场景的多源图像融合研究

The management and surveillance of people at certain public places become an important issue with serious consequences of human safety. Real-time monitoring, estimation and control of crowd appear to be essential. In recent years, a common weakness of surveillance systems is their inability to handle high crowded scenes. As the density of people in the scene increases, a significant degradation in the performance in terms of object detection, tracking, and event detection, is usually observed. This inability to deal with crowded scenes represents a significant problem given that many public areas are commonly densely populated. Based on the analysis mentioned above, in this project, we use the visual image and infrared image as the source images, thus more details of the scenes can be preserved. The research content includes crowd density evaluation and crowd flow analysis: 1) based on the correlation analysis theory of the data mining, feature vector of the high density scene can be constructed, which can be served as input value for regression algorithms. 2) In order to reduce the complexity the algorithm, graph-based semi-supervision learning and a Single hidden Layer Feed forward neural Networks would be introduced in our projects. 3）based on the multi-source image fusion feature vector of high-density population, propose segmentation algorithm of the sports groups. Innovation: firstly,multi-source image fusion algorithm based on a different level of granularity; secondly, optimize the feature vector-based semi-supervised learning to present accurate, rapid movement of the crowd segmentation algorithm.The aim of our project is to achieve precision density evaluation and flow analysis, so as to realize the auto-monitoring and obtain a qualitative output corresponding to the occurrence of an alarm condition. The research has a comprehensive application, and important academic value and social value.

本项目主要研究多光谱、多视角图像的特征融合问题。在该研究基础上，对高密度场景的人群实现自动、实时的密度估计及流量分析。主要内容包括：1）针对多源图像成像特性，基于数据挖掘的属性相关性分析来构建高密度人群场景特征向量的研究。2）基于纹理特征和半监督学习算法的人群密度估计研究。3）基于多源图像融合构建的高密度人群特征向量,提出运动群体分割算法。创新点：首先，提出基于不同粒度级别的多源图像融合算法；其次，基于半监督学习优化特征向量，提出准确、快速的运动人群分割算法。本项目的研究目标是对旅游景区、路口、广场、会展和商业中心和地铁等人群聚集的公共场所进行有效、实时人群密度精准估计和流量分析，从而实现人群的实现自动监控、自动报警。除此之外，该项研究成果对图像融合和运动群体的跟踪，具有非常重要的学术价值和社会价值。