Caracterização de funções de realidade aumentada usando computação de borda

Abstract

The growing use of immersive applications—such as augmented reality, virtual reality, and mixed reality—has led to increasing demand for more efficient computational resources capable of real-time data processing. While cloud computing can meet these demands, it often introduces high latency, negatively affecting the user experience. Edge computing emerges as a promising alternative by bringing computational resources closer to enduser devices, reducing latency, enhancing immersion, and enabling the deployment of such applications on mobile devices. Understanding the functional components of these applications and their performance profiles is essential for efficient offloading between mobile devices and edge servers. This work aims to characterize two core tasks in mobile augmented reality applications: Simultaneous Localization and Mapping (SLAM) and object detection. To this end, the MR-Leo prototype was used, integrating ORB-SLAM2 and incorporating a new object detection functionality based on the YOLO architecture. The research evaluates the performance of these tasks under different hardware configurations, considering execution on both CPUs and GPUs. The results show that although SLAM is computationally intensive, it performs acceptably on CPU-based architectures. In contrast, object detection requires massive parallelism for satisfactory performance and is heavily dependent on GPU usage. Based on the task characterization, a statistical workload model was developed to support the creation of workload generators capable of emulating the behavior of augmented reality applications under different computational architectures and scenarios.