基于视频监控数据的区域地理场景中人群特征分析

Intellgent perception and monitoring of crowd is in favor of enhance our ability to handle public health emergencies.Recently, video-based crowd feature analysis has gradually been in a hot research. However, these research methods analysis the corwd feature in image space, and have missed or ignored the relationship between the video analysis methods and the real geographical scene. The result is that the crowd feature extracted cannot be represented in geographical scenes by spatial location, spatial metric and spatial direction.Meanwhile, these research methods can only analysis crowd feature in video surveillance scenes, so the crowd feature cannot be analysised in surveillance blind zones. A framework which integrating video with GIS is proposed for crowd feature analysis. Key research points include the following: the bidirectional mapping model between video sequences to geographical scenes, video analysis of crowd feature coupling with geographical scenes, and the CA model of crowd feature estimation in blind area based on Ensemble Kalman filter. This is one beneficial exploration to crowd intellgent perception and monitoring problems in geographical scenes while with "Video-GIS" coupling analysis as the entry point. Through this research, the growing point will be generated in video analysis, GIS and other related areas, aiming to provide more solid theoretical foundation and methods reserve to the improvement of real-time monitoring ability and prediction and forewarning abilities for national public security.

密集人群的智能感知监控有利于提高突发公共事件的预防和处置能力。近年来，基于视频的人群特征分析已逐渐成为一大研究热点。由于这种方法是在图像空间中进行分析，忽视了人群特征与真实地理场景之间的联系，导致所提取出的人群特征无法在地理场景中用空间位置、空间度量与空间方位来描述。同时，这种方法只能获取视频监控区域内的人群特征信息，无法获取客观存在的视频监控盲区内的人群特征信息。本项目拟将视频与GIS进行有效集成，重点研究视频影像空间与地理场景空间双向映射模型、视频数据与地理场景耦合下的人群基本特征分析方法、基于集合卡尔曼滤波的视频监控盲区人群特征元胞自动机模拟方法。本项目是以"视频-GIS"耦合分析为切入点，研究地理场景中人群智能感知问题的一次有益探索。通过本研究，有助于在视频分析与GIS及相关领域形成新的学科生长点，有望为提高我国公共安全的实时监控、预警预报能力提供更加坚实的理论基础与方法。

密集人群的智能感知监控有利于提高突发公共事件的预防和处置能力。本项目以GIS空间关系、人工智能、摄影测量与计算机视觉等理论为基础，深入研究了视频空间与地理空间双向映射方法，构建了基于SIFT匹配的云台相机姿态变化检测方法、基于SIFT匹配的图像空间与地理空间自动映射方法，一方面解决了视频分析结果在地理空间中展现的问题，另一方面也为将地理空间约束引入视频分析奠定了基础。在此基础上，研究了基于视频图像的行人检测与跟踪方法，实现了基于带记忆的ViBe的运动目标提取、基于改进HOG算子的行人检测与基于粒子滤波的行人跟踪。同时，针对高密度人群，分别研究了基于支撑向量机的行人密度估计、基于光流场的行人运动模式分析，以及面向监控盲区的人群状态时空演化方法。本项目的研究以“视频-GIS”耦合分析为切入点，有效的解决了地理场景的行人检测/跟踪与高密度人群的统计特征与行为特征分析，所形成的原型系统也验证了本项目相关研究成果的有效性与实用性。