Search

Embedded AI rised in recent years, AI algorithm deployment platforms need strong computing power, but embedded systems focus on the balance between performance and cost. The contradiction has become a great challenge for the development of embedded AI. This paper analyzes the “ARM+FPGA” architecture and NEON register resources of the ZYNQ series chips introduced by Xilinx, and proposes a method to reasonably schedule the above resources to accelerate video preprocessing when it is used as an edge processing platform for intelligent monitoring terminals. This method makes the video intelligent analysis process of the monitoring terminal achieve real-time.

Edge intelligent computing devices are often deployed in some extreme environments, where the transmission network bandwidth is low or the network environment changes greatly. Therefore, the traditional queue scheduling algorithms cannot guarantee the QoS of edge intelligent computing. WF2Q+ allocates bandwidth according to a fixed weight, which causes real-time data flow delay to increase when the network is unstable. The dynamic perception scheduling strategy proposed in this paper is to dynamically change the weight of WF2Q+ by dynamically sensing the backlog length of the queue. At the same time, combined with the queue scheduling algorithm of PQ, this algorithm can prioritize the transmission of real-time data with certain fairness. In addition, the token bucket algorithm is used to limit the sending rate of the device and prevent network congestion caused by burst data injection into the network. After experimental simulation, the improved algorithm can achieve good results on the delay index.