Next Generation Video Surveillance for Environmental Monitoring and Protection

近年来，视频监控系统在家庭、办公室、学校和购物中心无处不在。由于它们具有远程查看、监控和发送警报的能力，它们也在环境保护领域崭露头角。与目前工业上广泛使用的传统压力传感器、质量体积计算和气体吸收光谱相比，使用视频图像有可能在早期阶段检测到逸散性泄漏。然而，在需要检测和测量液体/气体泄漏和烟雾排放等非刚性多晶物体的情况下，视频监控在环境监测和保护中的应用尚未得到很好的研究。最近有证据表明，石油从管道和泵站泄漏、温室气体排放、荒地火灾和烟雾警报表明，迫切需要一个高效和有效的系统来监测多态物体，但这些环境危害的易失性和动态性质使它们在技术上具有挑战性。此外，如今的监控系统由不同类型的数据采集设备（多模式）组成。不同的模态数据有其自身的优势，可以融合在一起进行综合数据分析，但是现有的技术，在过去的几十年里，对彩色图像和刚性物体有很大的希望，还不能满足对液体/气体泄漏或烟雾排放的多模态分析和非刚性多态物体分析的需求。另一个挑战是，分析这些数据的计算成本可能很高，并且需要实时功能。因此，有必要推进轻量、快速和高效的网络。提出的研究计划的长期愿景是创建一个基于视频监控的强大，高效和有效的监控系统，以监控和保护我们的物理环境。为了推进我的长期愿景，本研究在未来五年的短期目标是提出创新的方法来解决与多模态分析/非刚性多态对象分析以及网络规模和复杂性相关的几个问题。特别是，我的目标是推进背景建模和减法方案，提出基于学习的非刚性对象的有效表示，探索基于学习的非刚性对象3D重建技术，并研究上述深度学习网络的轻量级设计和快速训练策略。拟议的研究计划将为高素质人员（HQP）提供一个全面的培训平台，以进行图像处理、计算机视觉、视频监控和机器学习方面的创新研究。

Video surveillance systems have become ubiquitous in recent years in homes, offices, schools, and shopping centers. They are also emerging in environmental protection, given their ability to remotely view, monitor, and send alerts. Compared with the traditional pressure sensors, mass volume calculations, and gas absorption spectroscopy that are widely used by industry today, using video images has the potential to detect a fugitive leak at a very early stage. However, the application of video surveillance to environmental monitoring and protection has not yet been well studied in cases where non-rigid polymorphic objects such as liquid/gas leaks and smoke emissions need to be detected and measured. Recent evidence of oil spilling from pipelines and pump stations, greenhouse gas emissions, wildland fires and smoke advisories show an urgent need for an efficient and effective system to monitor polymorphic objects, but the fugitive and dynamic nature of these environmental hazards makes them technically challenging to monitor. Moreover, nowadays a surveillance system consists of different types of data acquisition devices (multimodal). Different modal data has its own advantages and can be fused together for integrated data analysis, but existing techniques, which have been promising for color images and rigid objects over the last decades, cannot yet meet the demands of multi-modality analysis and non-rigid polymorphic object analysis for liquid/gas leaks or smoke emissions. An additional challenge is that these data can be computationally expensive to analyze and require real-time functionality. Thus, there is a need to advance lightweight, fast and efficient networks. The long-term vision of the proposed research program is to create a powerful, efficient and effective monitoring system based on video surveillance to monitor and protect our physical environment. To advance my long-term vision, the short-term objective of this research over the next five years is to propose innovative approaches to solve several problems that are associated with multi-modality analysis/non-rigid polymorphic object analysis and network size and complexity. In particular, I aim to advance background modeling and subtraction schemes, propose learning-based effective representation of non-rigid objects, explore learning-based 3D reconstruction techniques for non-rigid objects, and investigate lightweight design and fast training strategies for the aforementioned deep learning networks. The proposed research program will provide a comprehensive training platform for highly qualified personnel (HQP) to perform innovative research on image processing, computer vision, video surveillance and machine learning.