面向无人机的卷积神经网络硬件加速方案设计OA

Efficient convolutional neural networks are difficult to adapt to small edge computing systems carried by UAV due to their large number of parameters and operations.This paper proposes a FPGA+ARM software-hardware heterogeneous scheme to offload the CNN product operation to FPGA hardware acceleration to achieve high-performance and low-energy target detection.The convolutional gas pedal is designed based on lightweight YOLOv4-tiny network on ZYNQ platform,using DMA to achieve data streaming between IP cores,and using FPGA parallel processing advantages to perform deep optimization through data multiplex-ing,pipelining and other operations.The experimental results show that the average accuracy of the model is 73%,the power con-sumption of the system is controlled within 3 W,the number of transmitted frames per second is FPS24.6,and the energy efficiency ratio reaches 8.14 GOP/W,which is a 5-fold enhancement compared to GPU.