Intelligent video surveillance for abnormal event detection

开发能够自动检测异常事件的智能视频监控系统需要解决具有挑战性的计算机视觉问题，如目标检测、跟踪和图像识别。近年来，这些问题已经通过越来越健壮的算法得到解决，在标准基准上产生了令人印象深刻的性能。然而，这一进展在现实世界的应用中并不令人印象深刻，在现实世界中，绝大多数视频监控系统仍然需要人工关注和人工干预。因此，我们认为，在计算机视觉基本问题的研究和智能视频监控应用的发展之间存在着巨大的差距。本研究旨在通过提高视频监控系统的智能性来缩小这一差距。我们所说的智能，是指通过自动提取重要信息和识别异常事件来帮助人类减少认知过载。因此，我们的长期目标是通过有效地利用计算机视觉方法和视觉获取技术来开发解决方案，以更好地理解监控场景。为了实现这一目标，我们采取了一种方法，从研究最先进的应用程序的局限性开始。这第一步对于发现优先问题和调查相关方法，以便开发强大的视频监控系统至关重要。遵循这一方法，我们提出了两个短期目标。根据该方案的基本轴心，我们的目标是：(1)通过优化利用深度学习和机器学习方法，提高目标和人的实时跟踪。更具体地说，我们的工作将集中在视觉对象跟踪(VOT)和人类骨骼跟踪，目标是提高跟踪精度和速度。在应用程序级别，我们将利用我们在跟踪问题上所做的工作来(2)检测和分析异常事件。我们将通过探索使用平移-倾斜-缩放(PTZ)摄像头来监控适度拥挤的场景，以及使用RGB-D摄像头来进行基于骨骼的行为分析，来分别解决室外和室内应用环境中的问题。我们的研究项目在几个方面是创新的，因为我们研究了计算机视觉方法中几乎没有探索过的异常行为(例如自杀企图)，并提出了以前工作中没有研究过的原创性方法。因此，我们希望我们的工作能为智能视频监控系统带来被高度引用的研究和实用技术。此外，该计划还专门为培训HQP量身定做。随着与计算机视觉和机器学习相关的公司的出现，相关的HQP将在加拿大各地获得高需求的专业知识。

Developing intelligent video surveillance systems that are able to automatically detect abnormal events requires solving challenging computer vision problems, such as object detection, tracking, and image recognition. During recent years, such problems have been addressed by increasingly more robust algorithms, producing impressive performances on standard benchmarks. However, this progress is not as impressive in real-world applications, where the vast majority of video surveillance systems still require human attention and manual intervention. For this reason, we argue that a significant gap exists between the work on fundamental problems of computer vision, and the development of intelligent video surveillance applications. The present research aims at narrowing this gap, by enhancing the intelligence of video surveillance systems. By intelligence, we mean to assist humans and reduce cognitive overload by extracting important information and identifying abnormal events automatically. Our long-term goal is thus to develop solutions to better understand monitored scenes, through the efficient exploitation of computer vision methods and visual acquisition technology. To reach this goal, we take an approach that starts by studying the limitations of state-of-the-art applications. This first step is crucial to detect priority problems and investigate relevant methodologies in order to develop robust video surveillance systems. Following this approach, we propose 2 short-term objectives. According to the fundamental axis of this proposal, we aim at (1) improving real-time object and person tracking, by making the optimal exploitation of deep learning and machine learning methods. More specifically, our work will be focused on visual object tracking (VOT) and human skeleton tracking, with the objective of increasing both tracking accuracy and speed. At the application level, we will take advantage of our work on the tracking problem for (2) detecting and analyzing abnormal events. We will address both outdoor and indoor application contexts, respectively, by exploring the use of a Pan-Tilt-Zoom (PTZ) camera to monitor moderately crowded scenes, and an RGB-D camera for skeleton-based behavior analysis. Our research program is innovative in several respects, as we study abnormal behaviors that are almost unexplored in a computer vision approach (e.g. suicide attempts), and we propose original methodologies that have not been investigated in previous works. We thus expect our work to produce highly cited research and practical techniques for intelligent video surveillance systems. Moreover, the program is specifically tailored for training HQPs. With the emergence of companies related to computer vision and machine learning, involved HQPs will acquire an expertise in high demand across Canada.