Compression of Video Streams for Transmission from the Edge to the Cloud for Video Analytics Applications

该项目是与加拿大一家专门从事人工智能的公司Nuvoola Inc.的研究合作。Nuvoola正在开发一种多因素实时认知分析引擎，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信息源映射为可操作的情报。该引擎的应用被考虑在医疗、安全、物流、零售、文化、金融、政府等领域。Nuvoola采用了一种独特的AI引擎设计战略，该战略与可用的无数边缘和云资源和服务无关。Nuvoola的AI引擎支持智能监控应用程序，在这些应用程序中，来自分布式摄像头网络的视频流被分析并提供可操作的情报。目前的实现是有限的，因为视频流中的重要时间内容丢失了，这限制了Nuvoola工具可以提供的功能。此外，利用深度神经网络对时间和空间信息进行融合，提高了人的识别和再识别等监控功能的准确性。然而，将原始视频流(包括空间和时间信息)传输到云是昂贵的，而且由于带宽限制变得不可行。这项研究活动将研究从Nuvoola的边缘设备到云的视频流传输的方法和算法，在云中，传输带宽和传输内容的质量之间的权衡得到优化。具体地，为了实现可接受的识别率，需要高质量的感兴趣区域(ROI)，例如人脸、对象等，而背景场景的质量并不重要。该研究项目的预期结果将允许通过弱传输通道将视频流从边缘设备发送到云服务，在弱传输通道中，视频分析和识别算法仍然有效，因为重要的ROI以足够的质量发送。

The project is a research collaboration with Nuvoola Inc.; a Canadian-based company specialized in Artificial Intelligence. Nuvoola is developing a multi-factor real-time 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 in areas such as healthcare, security, logistics, retail, culture, finance, government, etc. Nuvoola adopted a unique AI engine design strategy that is agnostic to the gazillions of edge and cloud resources and services available. The Nuvoola's AI engine supports intelligent surveillance applications where video streams from distributed camera networks are analyzed and provides actionable intelligence. The current implementation is limited since the important temporal content in the video streams are lost, and that limits the features that Nuvoola's tools can provide. Moreover, the accuracy of the surveillance functions such as person recognition and re-identification was proven to be boosted when spatial and temporal information are fused using deep neural networks. Nevertheless, transmitting the original video streams, including spatial and temporal information, to the cloud is expensive and becomes infeasible due to bandwidth restriction. This research activity will investigate methods and algorithms to enable transmission of video streams from the Nuvoola's edge device to the cloud, where a trade-off between the transmission bandwidth and the quality of the transmitted content is optimized. In particular, for achieving acceptable recognition rates, high quality of the Regions of Interest (ROI), e.g., faces, objects, etc., is required while the quality of the background scene is not important. The anticipated results from this research project will allow for sending video streams from the edge devices to the cloud services over weak transmission channels where video analytics and recognition algorithms are still functional since the important ROI are sent with sufficient quality.