Image Feature Extraction Research In Robot Visual Servoing System

The Visual Servoing of a robot makes use of the feedback of visual information,namely a computer was used to deal with the features extracted from the image in orderto control the position as well as the posture of the robot’s terminal control device, whichis an important means for a robot to obtain outer information as well as conductautonomous control and intelligent interaction. Visual signal processing, cameracalibration, the whole process of image feature extraction and robot control wereinvolved in the system of Visual Servoing. The present study focused on the thoroughstudy of camera calibration at an early stage as well as the extraction of image feature.Firstly, the related basic knowledge of the visual servo system was introduced,which included coordinate transformation, camera projection model, edge detection aswell as other related theories. Then, a study on camera calibration was conducted, and acamera calibration method, which based on improved genetic algorithms, was putforward. Making use of the genetic algorithms, the internal parameters of the camera wasglobally optimized via this method, and the local extremal optimization was conductedvia the steepest descent method. Thus, this method has the advantages of the needlessnessof initial estimates, high precision as well as strong noise immunity.Secondly, the principle of the sub pixel edge detection, which based on spatialmoment, was introduced in detail. Meanwhile, the use of spatial moment was extended tothe extraction of sub pixel corner features. The improved algorithm proposed in thisthesis firstly made use of Canny operator to test whether pixel is edge or not, then theextraction of sub pixel edge was conducted, and the number of the moment of templateoperation point involved was decreased and the operation speed was accelerated. Then,the spatial moment polynomial was set up to judge whether pixel point was in thevicinity of the corner point or not, and the extractions of sub pixel feature of the edgefeature as well as the angel characteristic were conducted respectively. An experimentalsimulation was adopted in the present study, and this method was proved to be practicaland saved large numbers of calculation. Finally, an improved sub pixel edge detection algorithm, which based onOrthogobal Fourier–Mellin Moments, was put forward. This algorithm made use of theOrthogobal Fourier–Mellin Moments’descriptiveness as well as noise resistance tosmall targets. Firstly, the initial edge location of the region of interest as well as Cannyoperator was conducted, and edge parametersθ, k and h were calculated via threemoment templates. Then, parameter l was obtained via geometric knowledge as wellas nonlinear equation, and the number of moment templates involved in the calculationwas reduced from 5 to 3, hence the operation time was shortened and the precision ofthe extraction as well as the noise immunity was improved.