Monocular Robot Target Recognition Based On Neural Network Positioning Research

Nowadays,with the rapid developing of science and technology,the process of industrialization obtains the good development environment. An increasing number of robots is applied in the factory,and a batch of epochal factory gradually appears such as "semi-automatic factory", "black lamp factory" and so on.The applications of robots, not only improve the quality of the products, accelerate the production efficiency, but also save the labor cost and keep 24 hours continuous production.Target identification is one of the key techniques of the robot. J st like Humans,before doing everything, we want to get the corresponding information.Human mainly obtains the information by their eyes,and the robot mainly gains the information by its cameras.Take industrial workpiece as research objects Robot has a very wide range of applications,such as welding robot, sorting robot,conveying robot,Room robot and workpiece recognition robot. This article mainly take industrial workpiece as research objects,and studies the target identification technology of robot, at the same time, also gives a basic research on target localization,finally designs the hardware circuit with DSP at the core that is used for image recognition.To identify and locate workpieces, the first thing to do is through camera obtains the digital image information of workpieces.Then preprocesses and extracts the characteristics of the workpiece.Finally sorts,identifies and locates the workpiece.By using the BP neural network and SVM network which can identify the workpiece,This article introduces two models which are applied to workpiece recognition.The main contents of this dissertation are expressed as follows:(1)The knowledge of image preprocessing to workpiece was studied,including graying, binarization, anti-binarization, closed operation, etc.To study the statistical characteristics of workpieces circumference,I studies the Canny,Sobel and other edge detection method to workpieces.(2)I study Hu invariant moment features and four kinds of statistical features,and verify by experiment that the combination of Hu invariant moment features and four kinds of statistical features has better recognition effect.(3)This article introduces several kinds of common image recognition method and the experiment that respectively uses BP neural network and SVM network to extract the feature parameters of workpiece.Then the results of experiment are analyzed combined with the experimental data.(4)Through reducing the dimension of images, I studies the location problem of image that its dimension has been reduced by the BP neural network.(5)Finally combining with DSP C6000 series processor, I design hardware circuit with TMS320DM642 at the core,including schematics design and PCB design.the arithmetic of this article runs on the basic of hardware circuit.(6)The whole contents of this study are summarized and prospected.