Intelligent surveillance for event detection

检测和通知用户感兴趣的事件是智能视频监控系统的关键功能。最重要的要求之一是实时监控系统必须：1)在真实世界的条件下连续自主地运行，以及2)可靠而不会使用户因无用的检测而过载。这是我们在拟议的项目中解决的两个具体问题，我们的目标是开发一种高效的视频监控系统，智能地检测并向用户报告感兴趣的事件。这是一个非常具有挑战性的问题，因为不受约束的场景的视觉特征是不稳定的，这主要是由于照明的巨大变化。另一方面，感兴趣事件的定义可能因用户而异。目前，大多数现有系统使用可见光摄像机，并使用基本的变化检测技术在持续且均匀照明的区域运行以检测简单事件(例如入侵)。最近的一些系统为用户提供了通过绘制感兴趣区域并设置诸如方向和速度的运动参数来定制检测的可能性。但是，这种定制仍然是一个复杂而耗时的过程，大多数用户不愿遵循。因此，我们的研究将集中在设计方法和算法来检测用户上下文中感兴趣的事件。我们方法的新颖性是双重的。第一种是使用深度图，这种深度图可以在不考虑光照条件的情况下获得，使用最新的具有成本效益的RGB-D传感器。因此，通过向标准颜色特征添加3D位置信息，可以构建更健壮的外观模型。第二点在于使用户能够对所报告事件的重要性进行评级。因此，通过根据给定用户过去的反馈调整警报，可以改进检测。该项目的目标包括：1)4D领域(3D空间和时间)中的动作和事件的模型，2)识别动作和事件的方法，以及3)从用户反馈中学习以供将来检测的方法。

Detecting and informing users about events of interest is the key function in intelligent video surveillance systems. Among the most important requirements is that the real-time surveillance system must: 1) operate continuously and autonomously in real-world conditions, and 2) be reliable without overloading the user with useless detection.These are the two specific problems that we are addressing in the proposed project, where we aim to develop an efficient video surveillance system that intelligently detects and reports events of interest to the user. This is a very challenging problem since the visual characteristics of an unconstrained scene are unstable, mainly due to large variations in illumination. On the other hand, the definition of an event of interest may vary between users. Currently, most of the existing systems use visible-light cameras and operate in constantly and evenly illuminated areas to detect simple events (e.g. intrusion), using basic change detection techniques. Some of the recent systems give the users the possibility to customize the detection by drawing interest zones and setting motion parameters such as the direction and velocity. But, this customization remains a complex and time-consuming process that most users are not willing to follow. Our investigation will therefore focus on designing methods and algorithms for detecting events of interest in the user`s context. The novelty of our approach is twofold. The first consists in using depth maps that can be obtained regardless of the illumination conditions, using recent cost effective RGB-D sensors. As a consequence, more robust appearance models can be constructed by adding 3D position information to standard color features. The second consists in enabling the user to rate the importance of a reported event. Therefore, detection would be improved by tuning the alerts for a given user based on his/ her past feedback. This project objectives include the development of: 1) models for actions and events in the 4D domain (3D space and time), 2) methods for recognizing actions and events, and 3) methods to learn from user feedback for future detection.