Real-time statistical anomaly detection in video

视频中的实时统计异常检测

• 批准号: 531338-2018
• 负责人: Gu, Hong
• 金额: $ 1.82万
• 依托单位: Dalhousie University
• 依托单位国家: 加拿大
• 项目类别: Engage Grants Program
• 财政年份: 2018
• 资助国家: 加拿大
• 起止时间: 2018-01-01 至 2019-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=652556
• 关键词: Real; time; statistical; anomaly; detection

EhEye is a surveillance recognition software company specializing in public safety and video surveillance by**using analytics and artificial intelligence. Surveillance video cameras are now widely used, however,**surveillance systems are often ineffective due to the need for experienced individuals' continuous attention to a**large number of videos. Automatic and fast detection of anomalous behaviour of individuals, such as fighting,**in a sequence of video frames offers new exciting possibilities for improving this effectiveness, especially for**large, critical infrastructure locations. This will provide a system which can optimize the capabilities of**existing security camera infrastructure and increase the security team's ability to respond to situations in real**time.**There are, however, a number of technical challenges in implementing such an automatic system. This project**will address the two key challenges - namely fast preprocessing of the data and accurate detection of anomalies**in video stream. First, the high dimensional data in each video frame needs to be reduced to the most relevant**low dimensional information needed for effective detection; the computation of this low dimensional**information needs to be performed in real time as video proceeds. Second, an accurate model needs to be**trained based on the extracted low dimensional information to estimate the probability of the fighting scenes**and conclude on it. The model or method for such prediction can be trained off-line but need to run in real time**when applied online.

EhEye是一家监控识别软件公司，专门从事公共安全和视频监控，** 使用分析和人工智能。监控摄像机现在被广泛使用，然而，由于需要有经验的个人持续关注大量视频，因此监控系统通常无效。自动和快速检测视频帧序列中的个人异常行为，例如打架，** 为提高这种有效性提供了新的令人兴奋的可能性，特别是对于 ** 大型关键基础设施位置。这将提供一个可以优化 ** 现有安全摄像头基础设施的功能并提高安全团队真实的 ** 实时响应情况的能力的系统。**然而，在实施这样的自动系统中存在许多技术挑战。该项目 ** 将解决两个关键挑战-即快速预处理数据和准确检测视频流中的异常 **。首先，需要将每个视频帧中的高维数据简化为有效检测所需的最相关的 ** 低维信息;随着视频的进行，需要在真实的时间中执行该低维 ** 信息的计算。第二，需要基于提取的低维信息 ** 训练精确的模型，以估计战斗场景的概率 ** 并对其进行推断。用于这种预测的模型或方法可以离线训练，但在线应用时需要在真实的时间 ** 中运行。