| dc.description.abstract | Effective disaster response depends on technologies that enable timely gathering and processing of data from different disaster sites in the disaster area. However, a large scale disaster like Hurricane Maria 2017 may totally destroy the infrastructure of the affected area, which makes both the Internet and cloud unavailable and the disaster response inefficient. To deal with this issue, Edge Computing and Communication (ECC), which performs data processing and data transmission at the edge of network, is applied to provide temporary computing and communication services to the first responders. By ECC, the first responders can process a large amount of sensing data at each disaster site and only send the processing result back to the Emergency Operation Center (EOC) via Disaster Response Networks (DRNs). However, the mobile devices carried by the first responders have limited computing resources, the wireless network connecting them is dynamic, and the applications used to analyze sensing data are computation-intensive and have diverse performance goals. Therefore, a lot of challenges exist for achieving high efficient ECC for disaster response.
In this dissertation, in order to address the aforementioned challenges, we propose an adaptive edge computing and communication framework for disaster response. This framework consists of a Distributed Mobile Stream Processing (DMSP) platform, a CNN-based multitask video processing system, and a user-customizable delay-tolerant routing protocol. Specially, the mobile stream processing platform allows first responders to perform computation-intensive stream processing on a cluster of mobile devices. It adopts feedback-based task configuration, resilient task assignment and adaptive stream grouping to deal with dynamic computing resources and network connectivity. The CNN-based multitask video processing system allows first responders to extract different Information of Interests (IoIs) from the on-body camera video stream by using different CNNs derived from the same base CNN. These CNNs can adaptively share different amount of common layers to trade off between the total computation cost and inference accuracy. Each of these CNNs can be adaptively divided into two separate parts to run on different mobile devices to meet the specific performance goals. The user-customizable delay-tolerant routing protocol enables first responders to send back different information obtained at each disaster site to EOC based on the specific Quality of Service (QoS) requirements. We evaluate the proposed framework through extensive real-world experiments and simulations, which demonstrate its effectiveness in enabling high efficient ECC for disaster response. | en |
