PERFORMANCE-SCORING DRIVEN MODEL SCALING AND SCHEDULING FOR EDGE VIDEO ANALYTICS SERVING

The growing demand for real-time, multi-task video analytics at the edge has encountered challenges in resource-constrained environments, including redundant computations across tasks and poor adaptability to dynamic workloads. In this paper, we propose a performance scoring-driven framework for model scaling and scheduling in edge video analytics. The framework consists of two core modules: (1) the model performance scoring module evaluates the model performance from four dimensions—video complexity, task correlation, model performance, and system resource utilization. (2) The model scaling and scheduling module then calculates a comprehensive score based on these four evaluation metrics. Aiming at maximizing the comprehensive score, this module employs the particle swarm optimization algorithm for model scheduling and system resource allocation, and selects the optimal combination of detection models based on the current model and system states. Experimental results demonstrate that our framework outperforms state-of-the-art baselines, achieving superior performance under dynamic edge workloads.