AdaSch: Load-Adaptive Scheduling for Accelerating Real-Time Edge Video Analysis With Diverse Inference Modes

Inference with spatio-temporal filtering plays a vital role in accelerating emerging edge video analysis. However, most existing solutions rely on a single inference mode, limiting their ability to cope with the heterogeneity and dynamic nature of edge systems. While combining multiple inference modes can enhance the real-time performance of video analysis, achieving efficient load scheduling across these modes remains challenging due to diverse device configurations and varying task requests. To address this important challenge, we proposeAdaSch, a novel load-adaptive scheduling framework designed to accelerate real-time edge video analysis using diverse inference modes. First, we design a new inference mode selection mechanism based on Region-of-Interest (RoI) features in different scenarios of video analysis. Next, we develop a Model Predictive Control (MPC)-based RoI scheduling method, which predicts continuous RoI sequences and obtains the optimal scheduling using a linear programming solver. Finally, we design a probabilistic sparse self-attention mechanism to capture server-side load patterns, enabling inference scheduling with optimal load balancing by solving a relaxed convex optimization problem. Extensive experiments on the real-world testbed and video datasets demonstrate the superiority ofAdaSch. Compared to state-of-the-art methods,AdaSchimproves accuracy and Service Level Objective (SLO) satisfaction rate by approximately 19.20% and 42.94%, respectively. Moreover,AdaSchexhibits excellent adaptiveness across various scenarios while maintaining minimal overheads.