Accelerating Real-Time Multi-View Edge Video Analytics with Spatial-Temporal Correlation

Advances in edge computing and video analytics are driving the widespread use of real-time IoT-based sensing and monitoring systems at the network edge. However, as camera clusters expand and video resolutions increase, meeting stringent response time requirements for video analytics is increasingly challenging due to limited computational resources and variable network bandwidth. To address these challenges, we propose a spatial-temporal feature-aware node selection scheme designed for multi-view edge video analytics. Our approach leverages spatial redundancy across video feeds to dynamically select specific processing nodes, while exploiting temporal correlations from prior frame detection to guide subsequent frame analysis, thereby reducing the computational load and enhancing system efficiency. Furthermore, we formulate an optimization problem for joint node selection and resource allocation, balancing the trade-off between end-to-end latency and inference accuracy. To tackle the complex coupling and temporal correlation, we propose a two-stage optimization scheme (PPO-LAG), that resolves the action dimensionality of Proximal Policy Optimization (PPO) through iteratively applying the Lagrange multiplier for interresource allocation. Experimental results on Wildtrack dataset show that our proposed scheme comparatively achieves higher inference accuracy while adaptively reducing the total latency.