Abstract
For public security purposes, distributed surveillance systems are widely deployed in key areas. These systems comprise visual sensors, edge computing boxes, and cloud servers. Resource scheduling algorithms are critical to ensure such systems' robustness and efficiency. They balance workloads and need to meet real-time monitoring and emergency response requirements. Existing works have primarily focused on optimizing Quality of Service (QoS), latency, and energy consumption in edge computing under resource constraints. However, the issue of task congestion due to insufficient physical resources has been rarely investigated. In this paper, we tackle the challenges posed by large workloads and limited resources in the context of surveillance with visual sensors. First, we introduce the concept of virtual nodes for managing resource shortages, referred to as virtual node-driven resource scheduling. Then, we propose a convex-objective integer linear programming (ILP) model based on this concept and demonstrate its efficiency. Additionally, we propose three alternative virtual node-driven scheduling algorithms, the extension of a random algorithm, a genetic algorithm, and a heuristic algorithm, respectively. These algorithms serve as benchmarks for comparison with the proposed ILP model. Experimental results show that all the scheduling algorithms can effectively address the challenge of offloading multiple priority tasks under resource constraints. Furthermore, the ILP model shows the best scheduling performance among them.