A Research Of Intelligent Recognition And Sorting Technology Of Specific Objects

In the face of the urgent needs of traditional industries to use artificial intelligence technology for industrial upgrading,robots are widely used in all walks of life.The current mainstream application is generally to replace manual labor for repetitive labor.This article aims to improve the intelligence of robot applications,and researches on intelligent recognition and sorting technology for specific tasks.It mainly studies three key technologies such as visual depth ranging and positioning technology,target detection and instance segmentation based on deep learning,and industrial robot control.and the development and verification of the corresponding subsystems are completed.In view of the dilemma faced by the environmental protection industry in China at the current stage of the public's weak waste classification awareness,the thesis designs a common rubbish intelligent recognition and robot sorting system,and uses the above research results to implement the system.Then,the system was tested in the garbage sorting and discarding condition to verify the feasibility of the technical solution.The main research contents of this thesis are as follows:1.The visual ranging and positioning technology is studied,and the instance segmentation technology is used to improve the traditional visual ranging method.The traditional visual ranging method is to add up the depth values of all pixels in the rectangular target box and calculate the average.Obviously,the target box contains a large number of non-target pixels.Therefore,this thesis first uses the mask obtained by the instance segmentation task to accurately segment the pixels contained in the target box.Then,the data processing method to filter outliers and partial edge values in the pixel values is used to achieve the improvement of ranging accuracy.Finally,according to the positioning principle of the RealSense depth camera,the SDK can be used to convert the position of the target in the image into three-dimensional coordinates in the camera coordinate system.2.An instance segmentation model algorithm based on Cascade Mask R-CNN is improved and implemented.The specific work is to use the GA-RNP network to replace the original RPN network,and connect and merge the mask branches corresponding to the three stages of the cascade network to gradually improve the segmentation effect.The experiments show that the improved model has improved performance.Finally,the model is used to complete the detection and recognition of objects in the picture and the pixellevel segmentation.3.the kinematics model of the UR3 robotic arm is constructed based on the D-H parameter representation method in the thesis,and it is very detailed to derive the mathematical models of its forward and reverse kinematics mathematical models.Then,the principle and theoretical model of traditional hand-eye calibration are analyzed in detail,and the automatic calibration experiment of hand-eye separation mode is completed using UR3 robotic arm.Its experimental results prove that the accuracy of the hand-eye calibration system fully meets the needs.4.This thesis designs and implements a garbage intelligent recognition and sorting system,it is integrated the above research results,and the construction of the hardware part of the system and the development of the software part are completed.The garbage classification scenario is simulated to test the system.It confirms the effectiveness of the technical solutions of the article.