Research On Object Recognition And Grab Position Based On Deep Learning

The recycling,classification and comprehensive utilization of solid waste in China are still at the initial stage,and there is no clear future planning and corresponding policies and measures.The comprehensive utilization rate of solid waste is low and the treatment method is backward.There is a lack of new equipment and new technology which are all problems encountered in the process of solid waste disposal in China.In the context of the era of Internet and big data,the breakthroughs brought by deep learning in recent years need not to be elaborated any further.The breakthroughs in voice application,face recognition,and human versus machine chess match are all derived from deep learning,which has become one of the most active research areas in the field of artificial intelligence.Therefore,applying deep learning to the classification of solid waste has become a research trend as well.This paper uses the convolutional neural network method in deep learning to solve the problem of classification and grabbing of solid waste and provides technical support for intelligent solid waste sorting system.This paper takes the intelligent sorting robot from Beijing Onky Intelligent Technology Company as the research platform.The main work is as follows:(1)Design and construction of the hardware part of the vision system.The overall architecture of the machine vision system was designed according to the needs of system.The hardware mainly includes industrial cameras,lenses,lasers and illumination sources.The selection and comparison of these components were made for the practical working environment.The working principle of the depth camera and the selection of the illumination method for color camera are explained.(2)Comparison of common target detection models.This paper selects the self-made solid waste special data set(mainly including seven categories of waste:wood,metal,brick,concrete,foamed concrete,paper and plastic)as the research object and explores the accuracy and speed of the Faster RCNN and YOLO with different shared convolution layers in solid waste detection and identification tasks.(3)Research on object grabbing and positioning.A rectangle grabbed by a robot is given by five-dimensional data:g={x,y,w,h,?},where(x,y)is the center of the rectangle,? is the direction of the rectangle relative to the horizontal axis,h is the height of the rectangle,and w is the width.First,a widely used convolutional network is used for grabbing prediction.This network has five convolutional layers and three fully connected layers.Then,a multi-model fusion method is used to input a RGB image and a depth image which converted into a 3-channel image into two independent VGG16 networks.The extracted feature values of the two networks are connected together and input into a convolutional neural network consisting of three fully connected layers,the convolutional neural network is used for grabbing prediction.