Collaborative Deep Learning-based Video Stream Face Recognition with Load Balancing Strategy via RBC and GPU Scheduling

Artificial Intelligence (AI) products have been widely integrated into human life, and these applications often involve multiple Deep Learning (DL) models, characterized by high computational demands and real-time performance requirements. We extend the GRA and GRA++ model based on E-CARGO for the development of the Group Role Assignment for Face Recognition (GRA4FR) model, which is based on Role-Based Collaboration (RBC) in cooperative services, to comprehensively manage the offloading strategy of terminal user tasks between terminal devices and Edge Computing (EC) servers to solve the joint communication and computing resource allocation problem. Furthermore, a two-level Load Balancing (LB) strategy is proposed: the first level targets EC servers using RBC-based methods, while the second level manages GPU scheduling for multi-GPU machines. Through real-world scenario simulations, we demonstrate that this dual-layered LB approach is crucial for maintaining system performance under dynamic workloads. Additionally, the mathematical formulation of task computational workload, as a key variable affecting system performance metrics, is further refined. Simulation results highlight the factors (the number of individuals on screen, the rate at which people pass by) as significant influences on system throughput and latency, underscoring the need for adaptive load balancing. Extensive numerical evaluations validate that our proposed method significantly enhances system efficiency, ensuring reliable performance in high-demand edge computing environments.