Network Resource optimization for Multi-View Streaming Mobile Augmented Reality

Mobile Augmented Reality (MAR) applications are demanding in computing and caching resources to support efficient amalgamation of Augmented Reality Objects (AROs) with the physical world in multiple video view streams. In this paper, the MAR service is decomposed and anchored at different edge nodes to enable efficient processing of popular view streams embedded with AROs. More specifically, we augment the notion of content popularity not only to objects but also video view streams and as such popular view streams are cached in a proactive manner, together with preferred AROs, in adjacent edge caching locations to ensure an acceptable level of user experience during different mobility events. To achieve that, a joint optimization problem considering mobility, service decomposition, and the balance between service delay and the preference of view streams and embedded AROs is proposed. To tackle the curse of dimensionality a nominal Long Short Term Memory (LSTM) neural network is proposed, which is trained offline with optimal solutions and provide real-time decision making during inference. Evidence from a wide set of numerical investigations shows that, the proposed set of schemes that provide service decomposition outperform nominal schemes which are oblivious of user mobility and the inherent multi modality of the MAR service.