A Comparison of Siamese Convolutional Neural Networks for Person Re-Identification in Video Surveillance**

视频监控中人员重新识别的连体卷积神经网络的比较**

• 批准号: 533701-2018
• 负责人: Granger, Eric
• 金额: $ 1.82万
• 依托单位: École de technologie supérieure
• 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=652318
• 关键词: Comparison; Siamese; Convolutional; Neural; Networks

Nuvoola Inc. is developing a multi-factor cognitive analytics engine that relies on multiple sources of information, and leverages artificial intelligence, cloud computing, and business rules to map these sources into actionable intelligence. Applications for this engine are considered for applications in healthcare, security, logistics, retail, finance, government, etc. A main component of the cognitive analytics engine is the sensor fusion module that combines sources of information from the different factors that are distributed over space and time., e.g., cameras, microphones, sensors, etc., to improve confidence on actions. ** In video surveillance, the person re-identification task is a use case of great interest to Nuvoola, where individuals must be recognized over time, across a distributed network of video cameras. However, the performance of techniques for person re-identification is typically poor in practice due to variations in camera viewpoints, capture conditions (e.g., pose, illumination, blur), background clutter and occlusion. Given the state-of-the-art accuracy achieved with deep learning on many challenging visual recognition problems, Nuvoola seeks to evaluate deep convolutional neural networks (CNNs) to design a spatial-temporal version of the sensor fusion module of their multi-factor cognitive analytics engine. ** For this project, the sensor fusion module combines image data acquired using video cameras, and some specialized algorithms will be investigated and developed for spatial-temporal recognition. Of a particular interest are Siamese CNNs that are suitable for accurate metric learning and recognition, potentially through domain adaptation on large-scale video datasets. Accordingly, the performance of several Siamese CNN architectures that allow for domain adaptation will also be evaluated in this project. The techniques developed and evaluated in this project are of great interest to Nuvoola (as a use case of the cognitive analytics engine), and to the computer vision and machine learning communities in general.** ********** ****************

Nuvoola Inc. 正在开发一种多因素认知分析引擎，该引擎依赖于多种信息源，并利用人工智能、云计算和业务规则将这些来源映射为可操作的情报。该引擎的应用被考虑用于医疗保健、安全、物流、零售、金融、政府等领域。认知分析引擎的主要组成部分是传感器融合模块，它结合了分布在空间和时间上的不同因素的信息源，例如摄像头、麦克风、传感器等，以提高行动的信心。 ** 在视频监控中，人员重新识别任务是 Nuvoola 非常感兴趣的一个用例，其中必须通过分布式摄像机网络随着时间的推移识别个人。然而，由于摄像机视点、捕捉条件（例如姿势、照明、模糊）、背景杂乱和遮挡的变化，人员重新识别技术的性能在实践中通常很差。鉴于深度学习在许多具有挑战性的视觉识别问题上实现了最先进的准确性，Nuvoola 寻求评估深度卷积神经网络 (CNN)，以设计其多因素认知分析引擎的传感器融合模块的时空版本。 ** 对于这个项目，传感器融合模块结合了使用摄像机获取的图像数据，并且将研究和开发一些专门的算法用于时空识别。特别令人感兴趣的是连体 CNN，它适用于精确的度量学习和识别，可能通过大规模视频数据集的域适应来实现。因此，该项目还将评估几种允许域适应的 Siamese CNN 架构的性能。 Nuvoola（作为认知分析引擎的用例）以及整个计算机视觉和机器学习社区对本项目中开发和评估的技术非常感兴趣。** ********** ****************