| With the rapid development of electronic,computer and artificial intelligence,the application of machine vision in all walks of life has become more extensive and mature.Among them,in the fields of automatic driving,augmented reality,and robot operation,vision devices and robots with object detection and pose estimation functions play a crucial role.For pose estimation tasks,object detection and localization are the premise and critical part of pose estimation.In the past ten years,the rapid development of machine vision,especially convolutional neural networks based on deep learning,has achieved good detection results in the tasks of object detection and pose estimation,but most of them have the problem that the detection accuracy and real-time performance cannot be perfectly balanced,and this dissertation is mainly oriented to the construction and comprehensive performance analysis of convolutional neural networks with high precision and high realtime detection performance of industrial robots for object detection and pose estimation.The preliminary research of this project has good application value and prospect.This paper compares and analyzes the data,methods and actual effects used before and after the improvement,and mainly does the following work:First,in order to alleviate the situation that the classical neural network model relies too much on a large number of images for predefined anchor box processing and ignores the rich semantic information of the deep features of the image itself,two SPP branch structures are added,a portable module mPAN is designed,and the reconvolution and upsampling network ReCUS is proposed.Considering the real-time and accuracy of the object detection model,this paper verifies the network structure of YOLOv4 and uses only some of the tricks of YOLOv4.After experimental comparative analysis,ReCUS convolutional neural network has better comprehensive performance.Second,in order to solve the problems of wrong or missed detection,based on the multi-model design of YOLOv5,the activation function,classification loss function and bounding box loss function used in classical neural networks are comprehensively analyzed,and the R-Mish activation function,M-BCE loss and S-CIoU loss function are proposed,based on which the Better-YOLO model is given.Experimental results show that the proposed methods have achieved different degrees of improvement on the PASCAL VOC 07+12 dataset.Third,the pose estimation model of MagicCubePose convolutional neural network is given.Based on the object detection network EfficientNet,the model retains its backbone network and extends it to build a 6D pose estimation module.The proposed pose estimation model in this paper is superior to other methods,achieving 97.5%ADD(-S)accuracy and 25.6 FPS.Forth,in order to solve the problem of single task of object detection model built based on COCO,PASCAL VOC and other public datasets,object detection database is built,including industry,campus,agriculture(pests),environmental protection and subarachnoid hemorrhage in medical images,multi-scene data,and a complete automatic labeling tool is designed to achieve 20-1000 classes of labeling tasks,and can realize common data formats such as txt,json,xml are converted to meet the training needs of different models. |
