| With the continuous development of sensor technology,more and more highprecision raw data are obtained by people.How to obtain higher-level information from complex and diverse data has become a hot topic under research.Benefiting from the substantial improvement of computer computing performance,the performance of machine learning methods represented by convolutional neural networks has improved by leaps and bounds,and its performance has surpassed traditional target detection algorithms in the same field.,target segmentation,etc.FPGA(Field Programmable Gate Array),as a semi-custom circuit,relies on programmable logic array to simulate any digital circuit function.Its low power consumption and parallel computing characteristics are in line with the operation process of convolutional neural network.This paper combines the regression-based target detection algorithm,deploys the convolutional neural network on the PYNQ development board for acceleration,and explores a high-performance and low-power neural network acceleration scheme on embedded devices.This paper first selects two lightweight neural networks MobileNetV2-SSD and YoloV4-Tiny suitable for deployment on edge devices,and uses the Pytorch deep learning framework to build the network model on the GPU.After training,the network model parameters,detection speed,Compared with indicators such as mAP.Secondly,the YoloV4-Tiny is deployed using the development board PYNQ-Z2 of the "ZYNQ+FPGA" architecture launched by Xilinx.The PYNQ board has the advantages of high performance,low power consumption and good flexibility.By dividing the tasks in the actual convolutional neural network deployment process,analyze the expansion resources and data reuse of the multi-dimensional convolutional layer,and use the HLS tool to complete the accelerated code writing,Verilog code generation and IP of the convolutional layer and pooling layer.Nuclear package export.Integrate the exported CNN acceleration IP core with the ZYNQ hard core in the Vivado environment and export the xsa platform file.Finally,use Vitis to write the software control process of the convolutional neural network,and complete the construction and testing of the CNN hardware acceleration system platform based on the PYNQ development board.The results show that the working clock of the PL side of the designed heterogeneous acceleration platform is 150Mhz,and the processing time of one image is 345ms.Under the condition that the performance is similar to that of the general-purpose processor Intel 8700K@3.7Ghz,the maximum power consumption is only 2.553 W.At the same time,the target detection prediction result on the KITTI dataset reaches a correct rate of 70.17%,the performance-to-power ratio has reached 135,which is higher than the CPU’s 2.6... |
