智慧城市中基於網路流量的可重組態影像串流服務

Abstract

智慧城市是利用人工智慧應用來解決許多現代都市問題的一個概念。同一時間，由於基於影像資料的許多人工智慧應用的蓬勃發展，相機裝置被大量地安裝在都市裡，再加上影像資料相對其他資料來說相對資料量大的天性，人工智慧應用需要在規模一邊快速成長的同時，一邊盡可能的保證影像資料品質。在今天智慧城市系統的規模下，基於代理者 (broker-based) 或是無需代理者 (brokerless) 的通訊模型各自面臨自己的挑戰，前者的系統中代理者可能成為系統運作的瓶頸，而後者則較難針對輕量化的相機裝置提供足夠的彈性以及可延展性。本篇論文提出一個可重組態的系統架構來解決兩種通訊模型遇到的挑戰，我們的目標是在他們遇到各自的運作瓶頸的時候提供更好的影像串流品質，這邊的串流品質我們更進一步聚焦到影像的幀率做為代表。我們提出的架構可以看作一個基於通訊模型的覆蓋網路，應用只需針對影像提出需求，該架構會動態偵測兩種模型各自可以提供的服務品質並用較佳的模型來做傳輸，該架構由一個負責偵測服務品質的品質偵測模組以及一個支援模式切換的可重組態方法組成，而且，當該架構採用無需代理者的通訊模型時，該架構可以釋放一些代理者的負載來讓其他使用代理者的應用得到更好的服務品質。實驗利用基於代理者的系統當作對照組與提出的架構進行比較，結果顯示提出的架構確實可以在代理者遇到頻寬瓶頸的時候提供給應用更多的影像幀數，並且可以讓代理者也提供更好的服務給其他應用。最後我們也討論了一些未來可能的研究方向。The smart city is a concept with AI-enabled applications introduced to mitigate challenges in modern urban areas. At the same time, many cameras are being installed inside the city for a wide range of applications. But AI applications face a significant challenge in ensuring data quality as the system scale rapidly grows. With the scale of the smart city today, common communication models like the broker-based model and the brokerless communication model both have their own challenges. The broker-based model has a performance bottleneck at the broker machine. The brokerless model can not provide good scalability and flexibility with lightweight camera devices.This thesis introduces a reconfigurable architecture to overcome the limitations of each model. Our goal is to provide better video data quality while one or both of the communication models meet their bandwidth limitations. The proposed architecture works as an overlay network on top of different communication models. The applications simply need to request the desired video data, then the proposed architecture will predict the optimal one of two communication models at runtime and deliver the data with it. The proposed architecture has a quality detector for predicting service quality and a reconfigurable method for switching communication models. Furthermore, the proposed architecture can also provide a positive side effect to the broker machine when the system adapts to the brokerless communication mode. We tested the proposed architecture by comparing it with the case that only adapts the broker-based model. The result indicates that the proposed architecture can provide better quality of video data at peak hours. And during the experiment, we also confirm the positive side effect that the proposed architecture can bring improvement to other traffic. And we also discuss some possible future directions of the proposed architecture.