Research On Object Recognition And Crawl Location Based On Machine Vision And Deep Learning

With the improvement of automatic level in industry and advancement of technology,the number of the robots keeps increasing,the application scene becomes more and more complex and the requirements of robots such as automation,intellectualization,accuracy,stability,flexibility also keep increasing.Due to the development of computer technology and artificial intelligence,the robots based on machine vision draw much attentions in the field of research and production.This paper researches on crucial technologies that refer to “object recognition and crawl location based on machine vision and deep learning”.The robot is divided into three parts: the hand,the eyes and the brain.The six degrees of freedom manipulator is regarded as the robot’s arm employed to crawl the object;the binocular machine vision are regarded as robot’s eyes employed to access external scenes;the deep learning technology is the robot’s brain employed to recognize and locate the object in scene.This paper researches on the main problems refer to hands,eyes and brain these three parts of the robot,the research contents contain the following aspects:(1)Researching on the camera calibration algorithm,aiming at the disadvantage of low extraction accuracyof target corner in the checkerboard calibration,we designed target recognition algorithms by using the plane circle target to calibrate the camera,and calibrate the monocular microscope using ZHANG’s camera calibration method.A feature point matching method that adds the epipolar constraints is proposed aiming at the problem that the limit constraint method for feature point matching tends to match wrongly when there are much more feature points.Finally,the calibration and match method need experiment validation.(2)Aiming at the problem of match between manipulator and visual coordinate system,a hand-eye calibration method based on the five-point circles target is proposed.Through the analysis of the hand-eye calibration experiment,and verify the effectiveness of this method.(3)Aiming at the control accuracy problem of manipulator,an improved multi variable PID neural network algorithm fused with multi innovation theory is proposed.In order to validate the convergence and the parameter estimation accuracy of the mehod,a MIPIDNN recognition system is designed.Its effectiveness is verified by the recognizing experiment of the Multi-input and multi-output nonlinear dynamic system.(4)A denoising method based on deep convolution neural network is proposed for the image denoising task in image preprocessing stage,and a nonlinear mapping between noise image and denoised image is composed by constructing a symmetric network consists of convolutional subnetworks and deconvolutional subnetworks.This method is proved to be effective by analyzing the denoise results of gaussian noise in different levels and other type of noise.(5)In order to make the manipulator to crawl the object in the field of view,we researched the SSD(Single Shot MultiBox Detector)object detection algorithm.Aiming at the problem that the SSD object detection algorithm has a low accuracy of location and recognition for small object,an improved SSD target detection algorithm based on feature pyramid is proposed to build and train the network model.Finally,we validate the network in the test data set in the large scale image detection competition,the average accuracy improved greatly and it has higher accuracy in recognition and location,validating that the improved method is effective.In conclusion,aiming at the object detection and crawl location of the intelligent manipulator system,the algorithm result proposed by this paper is obvious,and the research result has a certain theoretical significance and application value.