Research On FPGA-based Visual Angle Perception And Target Detection Device

With the rapid growth of our country’s power network,the spatial span of transmission and distribution lines has been greatly improved.The safety and stability of traditional manual and helicopter inspection need to be strengthened,and they can no longer meet higher efficiency requirements.Automatic patrol operation based on lightweight equipment carried by multi-rotor UAV can make full use of the UAV’s flexibility,low power consumption and low cost to provide more efficient,stable and safe patrol operation quality,and at the same time bring considerable benefits in terms of safety protection instead of manual operation.This thesis designs an FPGA-based embedded DCNN accelerator device to realize the perspective analysis and target recognition functions.Carry out meticulous inspection of transmission grid infrastructure and surrounding working environment,improve the inspection rate,reduce operational risks,provide efficient assurance for construction and inspection of power system,etc,further strengthen security protection of power system.In this thesis,PYNQ Z2 is chosen as the platform to implement ARM FPGA heterogeneous computing system FPGA is responsible for deploying DCNN accelerator and recognizing large scale convolution computing.A variety of optimization strategies for heterogeneous platform development is proposed to improve the speed of computation and detection rate,and the systematic design of software and hardware is completed.This research mainly includes three parts:ocular angle analysis through digital image processing,FPGA convolution neural network accelerator design and YOLOV3 target detection algorithm compression transplantation.Firstly,based on the DPU of Xilinx,the heterogeneous SoC design is carried out on the platform of PYNQ Z2,and the top layer of SoC is built to configure clock,terminal and constraint.The embedded Linux system is designed for PYNQ Z2 embedded development platform and DPU acceleration module.DPU driver is configured,root file system is generated,and SoC hardware platform is designed.The YOLOV3 network was trained in the Darknet framework,with targets including grid towers,pedestrians and large machinery.The YOLOV3 network model is converted to Caffe framework model,and the DPU IP structure is pruned,quantized and compiled to decrease the network parameters and output the light-weight binary network model.Finally,when the tower is recognized,Gaussian image denoting,Canny edge detection,Hough line transform and angle calculation are processed to get the visual angle estimation of the target tower and complete the visual angle perception task in the automatic patrol operation of the transmission network.The accuracy rate of target detection is 87.8%,system power consumption is 4.182W,the detection speed is 0.186S/MegaPixel,and the target detection speed,function embodiment,power consumption and cost are all improved greatly.