From Frames to Events: A Statistical Approach to Activity Analysis in Multi-Camera Systems

从帧到事件：多摄像机系统活动分析的统计方法波士顿大学，MA 02215与其他传感器不同，摄像机提供出色的分辨率，长可视范围，宽视场和低延迟，从而允许普遍的，广域的视觉监视。然而，今天部署的大多数网络摄像机都是简单的捕获/压缩/传输设备，最多支持基本的运动检测；所有高级处理都是高度集中的。这种集中式架构源于以人为中心的视觉分析以及有限的相机处理能力，并且无法扩展到大型多相机系统。目前，美国有超过3000万个监控摄像头在使用，每周产生40亿小时的视频片段，人工监控显然是不可持续的。需要一个自主的、分布式的、带宽高效的实时视频分析系统。这个项目朝着建立这样一个系统迈出了一步。其核心是用于活动发现和分析的新颖统计框架，该框架脱离了集中式模型，并利用了相机节点的处理能力。传统的活动分析是在对象层面上进行的，例如，对对象进行识别、跟踪和异常测试，而本项目正在开发的方法是在像素层面上进行活动分析。如果异常活动被可靠地识别出来，那么目标提取和跟踪就会集中在感兴趣的区域，因此相对简单，因为没有杂波。为了可靠地识别像素级异常，或更普遍的活动，使用了一种新的基于事件的视频表示，将视频分解为可用于统计学习的id样本。为了促进多摄像机协作，动态事件的几何不变性被利用，从而绕过了依赖于视角的三维几何相关的难题。

From Frames to Events: A Statistical Approach to Activity Analysis in Multi-Camera SystemsVenkatesh Saligrama and Janusz Konrad, Boston University, MA 02215Unlike other sensors, cameras provide excellent resolution, long viewing range,wide field of view and low latency thus permitting pervasive, wide-area visualsurveillance. However, most network cameras deployed today are simplecapture/compression/transmission devices, at most supporting rudimentary motiondetection; all higher-level processing is highly centralized. This centralizedarchitecture stems from human-centric visual analytics as well as limitedin-camera processing capacity, and is not scalable to large multi-camerasystems. With over 30 million surveillance cameras in use in the United Statestoday, that produce 4 billion hours of video footage per week, monitoring byhuman operators is obviously not sustainable. An autonomous, distributed,bandwidth-efficient, real-time video analytics system is needed.This project makes a step towards building such a system. At its core is anovel statistical framework for activity discovery and analysis that departsfrom the centralized model and leverages processing power of camera nodes.While traditional activity analysis operates at object level, e.g., objects areidentified, tracked, and tested for abnormality, methods under development inthis project employ activity analysis at pixel level. If the abnormal activityis reliably identified, then object extraction and tracking focus on region ofinterest and thus are relatively straightforward, on account of absence ofclutter. In order to reliably identify pixel-level abnormalities, or moregenerally activities, a novel event-based video representation is used thatdecomposes video into iid samples lending itself to the application ofstatistical learning. In order to facilitate multi-camera collaboration,geometric invariance of dynamic events is exploited thus bypassing thedifficult issues related to 3-D geometry dependent on viewing angles.