Abstract:In order to adapt to the increasingly widespread application of cameras in emerging fields such as smart city, intelligent transportation, and autonomous driving, video analytics have to respond with high accuracy and low latency. However, the high-precision analysis also brings a huge demand for computing resources, and cameras with limited resources cannot support the tasks. Edge computing can not only solve the problem of limited resources of local cameras, but also significantly reduce the transmission latency; hence, this scenario is considered in this work. By applying deep reinforcement learning methods, the analytics tasks on the specified cameras are dynamically selected to offload according to the current network conditions, to maximize the accuracy of the task over a period of time while satisfying the task response latency constraint. The simulation results show that the proposed method in this paper can guarantee the delay of tasks and maximize the task accuracy.