Research On Improved SSD Plane State Scattered Small Parts Sorting

At present,the industrial robots widely used in the production line can only perform repetitive actions.In essence,they belong to the teaching manipulator.They do not have the ability of independent perception and decision-making,and can not adapt to the grabbing and sorting of scattered and disordered parts.The combination of existing industrial robots and machine vision technology can solve the above problems and make the sorting robot have higher autonomy and flexibility.This paper systematically studies the target detection algorithm required for the sorting of planar scattered small parts based on depth learning algorithm,the depth estimation and attitude angle regression algorithm of slender parts,and the construction and control method of gantry three-axis manipulator.In terms of algorithm,this paper completes the development tasks of algorithm design,training,deployment and test evaluation,and realizes the functions of type recognition,position regression,depth estimation and attitude angle regression of planar scattered small parts.In the design and implementation of grasping actuator,this paper completes the hardware construction and control system design of three-axis gantry manipulator,and realizes the hand eye calibration and motion control of manipulator.The main research contents of this paper are as follows:(1)Design of target part existence judgment algorithm based on perceptual hash algorithm: calculate the similarity coefficient of two pictures through perceptual hash algorithm,and then judge whether there are parts to be sorted in the workbench.(2)Optimization design of target detection algorithm based on SSD: take SSD algorithm as the backbone and introduce it into the skills of algorithm detection head simplification,inner convolution operator and feature pyramid according to the specific application scenario of small part sorting,so as to improve the applicability of the target detection algorithm to this data set and achieve the purpose of reducing parameters and increasing efficiency.(3)Design of depth estimation algorithm for slender parts: according to the characteristics that different attitudes corresponding to different depths have great differences in imaging in industrial cameras,an image classification algorithm based on depth convolution neural network is designed to recognize different attitudes of such parts to achieve the purpose of depth estimation for slender parts.(4)Design of attitude angle regression algorithm of slender parts: design and implement the attitude angle regression algorithm of slender parts based on traditional image processing algorithm,so as to detect the attitude angle of slender parts on the workbench.(5)Manipulator construction and hand eye calibration: build a three-axis gantry manipulator based on the stepping slide,use the nine point calibration method to calibrate the industrial camera and manipulator,and obtain the coordinate transformation matrix from the pixel coordinates to the workbench plane.