On-line Estimation Method Of Image Jacobian Matrix In Robot Visual Servoing System

Robot visual servoing is an important research area in the field of robotics. The image Jacobian matrix, which as a local and linear approximation to the nonlinear mapping relationship between the changes of image feature and the changes of robot motion, is always required in image-based visual servoing. Therefore an exact image-based visual servoing control relies on the accuracy of the image Jacobian matrix, which needs a precise knowledge to the camera models and kinematic of robotic system. However, accurate calibration is hardly realizable in real applications. Therefore, uncalibrated visual servoing has gradually become a research focus in recent years.Uncalibrated visual servoing can avoid complex and time-consuming system calibration process of the traditional robot visual servoing based on calibrated techniques, therefore it has important theoretical significance and broad prospect for industrial applications. Based on the summary of the development of robot uncalibrated visual servoing, this paper studied the image Jacobian matrix on-line estimation methods in robot uncalibrated visual servoing from two aspects of filter method and nonlinear visual mapping model method. The main work in this paper can be summarized as follows:(1) To overcome the shortcomings of traditional image Jacobian matrix on-line estimation methods in dealing with nonlinear problems, an on-line estimation of image Jacobian matrix method based on the square-root unscented Kalman filter is proposed. In this method, a state vector is formed from the elements of a image Jacobian matrix, and the problem is converted into one of system state estimations, then a square-root unscented Kalman filter suitable for nonlinear systems is utilized for estimation of system state, thus the on-line estimation of the image Jacobian matrix is realized and the complex system calibration process can be avoided.(2) An on-line estimation of image Jacobian matrix method based on the online support vector regression is proposed. Due to the conventional intelligent learning methods based on neural network and support vector regression are required to be learned offline in advance, and retraining will be needed if the model is changed, therefore an on-line estimation of image Jacobian matrix method based on the online support vector regression is proposed to solve this problem. In this method, an online support vector regression suitable for fitting nonlinear mapping model is utilized for continuous learning the complex nonlinear mapping relationship between the image feature space and the robot joint space, thus the on-line estimation of the image Jacobian matrix is realized and the complex calculation of retraining can be avoided.(3) The experimental platform of the robot visual servo system is set up. The hardware system is constructed based on MOTOMAN-MH12 industrial robot, Smartek industrial camera and host control computer; the integrated software platform is developed based on MFC under Visual Studio 2010 development environment in the Windows system, and it includes image processing, image feature extraction, robot motion control, uncalibrated visual servoing algorithm and human-machine interface.Simulation and experiments for every kind of image Jacobian matrix on-line estimation method are carried out to verify the efficiency and validity.