Investigating resource efficient algorithms for monitoring in video surveillance

本项目旨在设计将在原型智能相机中实现的视频分析算法。它们是：1)检测未经授权的人/入侵和2)检测住宅。尽管科学文献中存在一些执行此类任务的方法，但它们中的大多数要么不是非常健壮，要么不适合在资源（处理器和内存）更有限的嵌入式平台上运行。我们的合作伙伴公司希望在嵌入式系统上实时执行这些任务。因此，我们的研究将集中在研究方法和算法上，这些方法和算法可以在智能摄像机中有效地实现，同时又具有鲁棒性，用于检测入侵和居住在监控视频中。我们将研究如何使用子采样来处理每帧更少的像素，但在背景减法算法中检测感兴趣的对象。我们还将研究后台更新机制，以选择或仅适应最有用的机制。此外，我们将设计一种运动历史方法来累积检测到的前景检测。决策将基于使用机器学习技术的积累图。

This project aims at designing video analytic algorithms that will be implemented in a prototype smart camera.They are: 1) Detection of an unauthorized person/intrusion and 2) Detection of dwelling. Although, severalmethods exist in the scientific literature for performing such tasks, most of them are either not very robust ornot adapted for running on a embedded platform with more limited resources (processor and memory). Ourpartner company wishes to perform on an embedded system these tasks in real-time. Our investigations willtherefore focus on researching methods and algorithms that can be implemented efficiently in a smart camera,yet that are robust, for detecting intrusion and dwelling in surveillance videos.We will investigate how subsampling can be used to process fewer pixels per frame, but yet detect the objectsof interest in background subtraction algorithms. We will also investigate the background update mechanismsto select or adapt only the most useful ones. Furthermore, we will design a motion history approach toaccumulate the detected foreground detections. Decisions will be made based on the accumulation map usingmachine learning techniques.