An optimized environment-adaptive computation offloading strategy for real-time cross-camera task in edge computing networks

With the large-scale establishment of cross-camera networks, edge computing plays an important role in real-time tasks with its abundant edge resources and flexible task offloading strategy. Conventional studies usually utilize cross-camera network topology and real-time task status to generate subtask offloading strategies. However, most existed approaches focus on utilizing static environment information to generate a fixed offloading strategy for single-target optimization, while dynamic environment information and joint optimization objectives are often ignored. In this paper, we model the computing process of cross-camera tasks as a Markov Decision Process (MDP) integrating spatiotemporal correlation, to make full use of the dynamic environment information in the edge computing network. In addition, to achieve multi-objective optimization of cross-camera tasks, this paper develops a joint Q learning equation that integrates multiple utility indicators and proposes a Deep Spatio-Temporal Q Learning (Deep-STQL) algorithm to solve the equation. Based on the camera frame rate and cross-camera task frame rate, a large number of experimental data show that our proposed Deep-STQL algorithm has significantly improved the convergence, hit rate, average processing delay, drop rate of subtask and computing load of real-time cross-camera tasks compared with the baselines.