FutureView: Predictive Video Partitioning for Distributed Low-Cost Mobile AR Analytics

Augmented Reality (AR) applications often struggle with compute-intensive video analytics, leading to high processing delay and significant energy drain on mobile devices. To overcome this, we introduce FutureView, a novel framework for cost-efficient distributed AR analytics. FutureView uniquely partitions each video frame into a Region of Display (RoD), requiring immediate processing for rendering, and a predictive Region of Processing (RoP), analyzed with greater delay tolerance for potential future viewpoints. We formulate the workload distribution across device, edge, and cloud resources as a constrained optimization problem, minimizing overall cost while preserving recognition accuracy. Furthermore, FutureView employs an online algorithm that dynamically adjusts the RoP size based on historical utilization and predicted viewport changes, maximizing resource efficiency. Extensive real-world testbed evaluations demonstrate that FutureView achieves a 26.9%–82.3% cost reduction across different baseline approaches, while maintaining high precision and exhibiting robustness to prediction noise and system variations, establishing its practical viability for resource-constrained AR applications.