基於行動擴增實境應用的本地端快取管理技術

Abstract

行動計算在近十年來進展卓著。其中一項新興的行動計算應用是透過行動裝置 支持的擴增實境(Augmented Reality, AR)服務，或稱為行動擴增實境(Mobile Augmented Reality, MAR)。例如:使用智慧型手機上的攝影鏡頭擷取使用者周圍環境的景象，並且 根據景象中的擴增實境目標物件適當地疊加虛擬延伸資訊於目標物件上，然後再一起 顯示於手機螢幕中。因為這樣的特性，行動擴增實境在行動商務與適地性服務等方面 引起了廣泛的興趣。然而，擴增實境是具有高度計算需求與通訊需求的服務。在移動 設備上執行的行動擴增實境服務受限於移動設備本地端儲存空間、計算能力與反應延 遲以及有限的無線通訊頻寬等因素。擴增實境的高度計算需求來自於目標物件可能需 要和資料庫中大量的特徵資料進行比對辨識。擴增實境的高度通訊需求來自於目標物 件可能需要被持續地傳送至伺服器端並且由伺服器重複下載相同的虛擬延伸資訊。因 此我們認為可以透過尋找物件相依性使得遠端處理的工作可以被移至本地端執行。本 論文研究本地端快取管理議題使得在行動裝置上的計算與通訊成本可以被限制在容許 範圍中並且平衡本地端與遠端的工作負載。本文設計了基於空間相依性與時間相依性 的快取管理策略使得本地端快取由遠端伺服器預取高優先權的物件，並且置換低優先 權的物件。本論文透過模擬實驗的三項指標:反應時間、遠端訪問次數與快取命中率來 驗證本文快取管理策略的有效性。最後我們也同時在Android平台上實作了一個雛型系 統驗證系統效能。

Mobile computing has made significant progress in the past decade. An emerging application in this field is to support Augmented Reality (AR) via mobile devices, or known as Mobile Augmented Reality (MAR). The real-life views from the camera of a smartphone can be overlayed with extended information from the virtual world according to the AR target objects. Thus, MAR has triggered strong interests in mobile e-commerce, location-based service, etc. However, MAR, when run on a mobile device, is usually constrained by the local storage and computing power of the device and the latency and bandwidth of the underling wireless channel. The computation demand is high because a target object needs to be compared to thousands or millions of images in the database. The communication demand is high because potential objects need to be continuously transmitted to the server and the augmented information needs to be downloaded from the server. It is thus desirable to explore locality by bringing some of the remote processing work to the local device. This thesis investigates the Local Cache Management (LCM) problem to manage the computation and communication costs of a mobile device within in an acceptable level and harmonize the local and remote workloads. We present strategies to pre-fetch higher-priority objects from the server and replace lower-priority objects in the local cache based on access temporal and spatial locality. We verify the effectiveness of our strategies in term of cache hit ratio, response latency, and remote requests via thorough simulations. We also show our lab prototype of LCM on Android to verify its feasibility.