Research And Design Of Target Detection System Based On Heterogeneous Embedded System

With the widespread popularization of 5G communication technology,traditional target detection systems have gradually been unable to meet the needs of various industries that are now entering intelligence.Consumers have put forward higher requirements for traditional target detection systems that relied on machine learning in the past,and related manufacturing companies have also begun to transition to intelligence and low power consumption.Among them,how to combine image enhancement technology,target detection technology and neural network to transplant to a portable embedded platform has become a hot topic for developers today.This subject designs a target detection system based on heterogeneous processors,which can efficiently and accurately identify target objects in a specific environment where some target objects are dense.This article first designs the overall target detection system.In terms of hardware design,the multi-core heterogeneous ARM processor RK3399 is selected as the core processor,the image acquisition module uses the CAM1320 module,and peripheral auxiliary circuits are built to complete the hardware design;software Part of the introduction of convolutional neural networks.Today's mainstream target detection algorithms extract feature information from the pixel matrix of the image through convolution and pooling processing,and perform category prediction on the image in the fully connected layer.And analyze the network structure,bounding box and loss function in the typical representative algorithm YOLOv3.By comparing the performance indicators of YOLOv3 and other target detection algorithms,YOLOv3 was finally selected as the basis for the improvement of the algorithm in this paper.Afterwards,the software part of the system's target detection is described as a whole.Firstly,the HSI-based image enhancement algorithm is introduced.The feature information and color information of the image are processed by multi-scale Retinex and color stretching,respectively,and the enhanced image is used as the input of the target detection algorithm.Then,in view of the target missed detection and detection classification concentration phenomenon shown by YOLOv3 in a specific detection environment,the YOLOv3 algorithm is optimized.On the one hand,the SE unit is introduced to enhance the features,and on the other hand,the original three-scale output is optimized as Four-scale output for dense environments.Finally,after the cross-compilation and debugging environment of the target detection system platform is built,the optimized neural network model is transplanted to the embedded platform,and the target detection system platform is tested.Through the various parameters of the image and the accuracy of target detection And other indicators are analyzed,and experimental conclusions are drawn,which prove that the target detection system can effectively enhance the image,and its recognition efficiency reaches 92.56%,and the system meets the high accuracy requirements of actual scenes.