Abstract : real-time human action recognition and behavior monitoring within video streams impose significant computational strains on backend server infrastructures. Traditional distributed system load balancers assign dynamic incoming media tasks based exclusively on infrastructure-side metrics like CPU utilization or memory bandwidth, completely omitting request-specific computational requirements. This mismatch results in suboptimal task allocation, frame drops, and execution latencies when multi-scale convolutional operations or dense optical flow models are triggered unpredictably. To resolve this bottleneck, this paper introduces a novel Request-Aware Fuzzy Load Balancing (RAFLB) framework. The proposed paradigm establishes an adaptive, two-phase scheduling ecosystem. First, high-throughput video streams are frame-decomposed and pre-processed using spatial-temporal filtering kernels and Lucas-Kanade optical flow equations to extract intrinsic stream metadata (resolution, frame rate, structural intensity). Second, a multi-input Mamdani Fuzzy Inference Engine computes real-time routing priorities by simultaneously processing the localized Request Weight (RQ) alongside Server Busy (SB) telemetry. Experimental simulations show that RAFLB drastically reduces structural frames latency by up to 34% and prevents cluster choke points compared to conventional round-robin and resource-only load balancers.
Shaxboz et al. (2026) studied this question.