Methods and algorithms for multi-camera visual surveillance and monitoring

该研究计划的目标是为智能监控系统的开发设计新的方法和算法。该研究将侧重于构建下一代视频监控系统所需的必要组件，该系统能够提取有关正在发生的事件的高级信息，并使用户更容易访问此类系统收集的信息。重点将放在可以集成到实时监控系统的解决方案上。 所寻求的解决方案将执行基于对象的视频分析，以检测、跟踪和重新识别场景中的特定对象，并识别它们的动作。我们的研究计划不仅仅是开发这些单独的模块，还旨在将这四个视觉组件无缝集成到实时视频监控系统中。为了验证，应用和扩展这项研究，我们的研究将在一个先进的监控系统的背景下进行，用于跟踪顾客在商店和餐馆的行为。 目标检测方法的显著进步使其成为视频中目标提取的良好候选者。在视频监控的背景下，这些可以被设计为在非常低的误报率下操作，即使这会导致重要的遗漏率。使用视频使我们能够恢复在以前的帧中丢失的对象。当需要时，可以从视觉跟踪器获得检测之间的对象位置。在多个相机上的长期跟踪可以通过重新识别模块来实现，该重新识别模块可以重新检测先前观察到的移动对象。最后，完整理解场景动态的最后一个基本要素是检测场景中演员所采取的动作。这四个组件代表了高级视频分析系统的基本构建块。 我们预计，拟议的研究计划将导致新的专业视频分析系统的开发。基于我们目前的合作，我们专门针对客户跟踪应用，或更广泛的公共场所事件检测和行为分析。虽然这项研究的大部分重点是对固定摄像机拍摄的视频的解释，但我们也打算考虑移动摄像机（如运动摄像机和可穿戴摄像机）的情况。

The objective of this research program is to devise new methods and algorithms for the development of intelligent surveillance systems. The research will focus on the necessary components required to build the next generation of video monitoring systems capable of extracting high-level information on occurring events and give users an easier access to the information collected by such systems. The emphasis will be on solutions that can be integrated to real-time surveillance systems. The solutions sought will perform object-based video analytics for the detection, the tracking and re-identification of specific objects in the scene and the recognition of their actions. More than the development of these individual modules, our research program aims at the seamless integration of these four vision components into real-time video surveillance and monitoring systems. In order to validate, apply and expand this research, our research will be performed in the context of an advanced surveillance system for tracking the behavior of customers in store and restaurants. The remarkable advances in object detection methods make them good candidates for object extraction in videos. In the context of video monitoring, these ones can be designed to operate at a very low false positive rate even if this leads to an important miss rate. Working with videos allow us to recover objects that were missed in previous frames. Object position in-between detections can be obtained, when required, from a visual tracker. Longer term tracking, over multiple cameras, can be achieved through a re-identification module that can re-detect a previously observed moving object. Finally, the last essential ingredient for a complete understanding of a scene's dynamic is the detection of the actions undertaken by the actors of the scene. These four components represent the fundamental building blocks of an advanced video analytics system. We foresee that the proposed research program will lead to the development of new specialized video analytics systems. Based on our current collaborations, we specifically target applications in customer tracking, or more generally event detection and behaviour analysis in public places. Although most of this research has a focus on the interpretation of videos captured by fixed cameras, we intent to also consider the case of moving cameras such as action cameras and wearable cameras.