III: Small: Distributed Semantic Information Processing Applied to Camera Sensor Networks

在许多应用中，传感器网络可以用来监控大范围的地理区域。这通常会产生大量数据，需要对这些数据进行关联、汇总和分类，以便对所监视的现象进行语义上有意义的描述。该项目的长期指导愿景是一个分布式网络，它可以在很长一段时间内以可扩展的方式自主执行此分析。作为一个具体的应用，本研究的重点是具有静态节点或机器人代理一部分的智能摄像机网络。计划中的工作将带来更高效、更准确和更具弹性的系统。开发的算法将得到广泛的应用，包括在安全(连续实时检测可疑个人)和物联网方面。作为更广泛影响的一部分，该项目将制作教育材料，向K12受众解释该项目的科学成果。这项研究将引入一个新的框架，用于在智能摄像机网络中一致、高效和可靠地提取语义信息。该项目将围绕三个支柱：(I)解决一致的多摄像机特征匹配问题：找出在一个节点处看到的特征是否被其他摄像机节点看到，如果是，则使用一种分散散列的变体来找到最一致的对应关系；(Ii)通过特征聚类，使用用于计算k-medoid的分布式算法，自主地发现新出现的模式(例如，来自不同个体的图像)；(Iii)采用新的分布式决策树推理范例，允许网络对世界做出一致的决策(例如，个人的身份)，同时将计算分散到网络上。该项目网站将提供对工作成果的访问，包括论文、可下载的代码和教育材料。

In many applications, sensor networks can be used to monitor large geographical regions. This typically produces large quantities of data that need to be associated, summarized and classified in order to arrive to a semantically meaningful descriptions of the phenomena being monitored. The long-term guiding vision of this project is a distributed network that can perform this analysis autonomously, over long periods of times, and in a scalable way. As a concrete application, this research focuses on smart camera networks with nodes that are either static or part of robotic agents. The planned work will result in systems that are more efficient, accurate, and resilient. The algorithms developed will find wide applications, including in security (continuously detecting suspicious individuals in real time) and the Internet of Things. As part of the broader impacts, the project will produce educational material to explain the scientific results of the project to a K12 audience.This research will introduce a novel framework for consistently, efficiently and reliably extracting semantic information across a network of smart cameras. The project will revolve around three pillars: (i) solving the consistent multi-camera feature matching problem: find if a feature seen at a node has been seen by other camera nodes, and if yes, find the most consistent correspondences using a variation of decentralized hashing; (ii) autonomously finding emerging patterns (e.g., images from different individuals) through feature clustering, using a distributed algorithm for computing k-medoids; (iii) employing a new distributed decision tree inference paradigm that allows the network to take consistent decisions about the world (for example, the identity of an individual) while spreading computations across the network. The project website will provide access to the results of the work, including papers, downloadable code, and educational material.