Research On The Key Technologies Of Feature Selection Image Based Visual Servoing

Visual servoing system by real time using visual information to realize the feedback control of the robot,can effectively improve the system application and design flexibility and intelligence level,is an important research direction in the field of intelligent robot.It involves multidisciplinary integration of computer vision,robotics technology and control theory and so on,and has always been the focus and difficult problem in robot research.According to the feedback information,visual servoing can be divided into two categories: position based visual servoing(PBVS)and image based visual servoing(IBVS).Compared with the PBVS,the IBVS method doesn't need 3D reconstruction,and does not depend on the observed target model.It is more robust to the system parameter errors and noise,and gets more extensive attention.Most of the existing methods are based on the image plane point feature to directly design the IBVS controller,the system has a high degree of non-linearity and coupling characteristics.The feedback signal is directly defined in the image plane,and the realization of the motion lacks the control and constraint of the camera's 3D task space,while there are potential problems such as Jacobian singularity and local minima.Feature selection is the basis and key part of IBVS.On the one hand,it can expand its application.On the one hand,it can overcome the potential problems existing in the traditional methods,reduce the system nonlinearity and obtain the appropriate motion decoupling,which can help to improve motion characteristics.In this paper,the robotics visual servoing of monocular eye-in-hand configuration is taken as the research objective.Based on feature selection,several key techniques of IBVS are researched,and the main aspects are as follows:Firstly,machine learning method is introduced for design of invariant visual features for IBVS on the basic of analyzing the kinematics of image moment features,an improved moment based IBVS method with learning features is proposed.Firstly,the kinematics of 2D moments and Centered Moments are researched based on the camera kinematics model and Green's theorem.The image moments and related Jacobian matrixes are respectively calculated for both target discrete image point set and the regional projection feature.Secondly,two virtual image moments are designed to be linearly related to the rotation angle of the X-axis and the Y-axis of the camera,kinematics of virtual features is analyzed and exhibits complete decoupling and linear characteristics,this also solves the problem of Jacobian singularity and makes it applicable to any shape of plane target.Based on the TRS(2D translation,2D rotation and scale transformation)invariant characteristics of the Hu invariant combination features,the "TRS invariant moments-rotation angles” recessive nonlinear mapping model is defined,then nonlinear support vector machine based on insensitive loss function is proposed to perform regression learning on the model.Finally,based on the characteristics of learning virtual image moments and the normalized image center of gravity coordinates,the normalized area feature and the direction angle feature,six independent image moments and corresponding feature Jacobian are obtained,and the task function is used to design the visual servoing controller.Secondly,based on the unified projection model,a decoupling visual servoing algorithm with virtual unit sphere projection is proposed.The projection feature of the target on the surface of the virtual unit sphere is calculated by using the unified projection inverse model.For the minimum point set(N = 3)and the N-point set(N> 3),with the rule of vector and matrix mixed derivation,a series of visual features invariant to rotational motions and an orthonormal basis pose features are designed,and the corresponding Jacobian matrix is deduced.According to the 3D translational motion control,invariant visual features from vector inner product operation,external product operation and mixed product operation are proposed;while according to the 3D rotational motion control,virtual point features are defined to design the orthogonal basis pose feature from arbitrary number of image points.For the case of N-point(N> 3)set,an invariant visual feature selection algorithm based on vector similarity analysis is proposed.Finally,a spherical IBVS controller partially decoupling translational and rotational motions is designed,applicable to the case using the point feature as the initial description of any 2D and 3D target.Thirdly,a multi-variable constrained visual servoing with spherical projection model is proposed.Nonlinear model predictive control is used to define spherical IBVS as an nonlinear constrained optimization problem.Based on the internal model control structure,the multivariable weighting objective function is defined,and the dynamic iterative equations of the 3D and 3D constrained variables is proposed to realize the multi-constraint active avoidance.Kinematics of feature vector combining mixed product invariant visual features and the orthogonal basis features are used to describe the behavior of dynamic system.First order Euler approximation of kinematic equation is proposed to predict the evolution of the visual features with respect to the camera velocity over a finite-prediction horizon.In order to make full use of the kinematic characteristics of spherical features,a local correction strategy for MPC optimization is proposed.Compared with solving the MPC optimization problem in the feasible region of the independent variable directly,the correction method takes precedence over speed constraints,and takes the proportional transformation of the constraint violation to obtain the modified control sequence.Fourthly,an uncalibrated visual predictive control algorithm based on discrete reference trajectory is proposed.Under the framework of visual predictive control,the robot motion control based on direct visual servoing kinematics is realized without any knowledge of robot model,camera model and target model,and the feasible domain constraints of the camera field of view,the joint angle of the robot and the joint velocities can be well handled.In order to achieve the prediction of the system behavior over a finite-prediction horizon,a prediction model based on the approximate kinematics is proposed by the composite Jacobian dynamic estimation.A Jacobian global estimation method combining KNN algorithm and RLS dynamic iteration is proposed to improve the effectiveness of prediction.Meanwhile,in order to improve the dynamic convergence characteristics of the system,a model-independent projection interpolation algorithm is proposed for the reference trajectory discretization design.Path planning in the image plane is realized by the feature affine transformation and similar matrix Eigendecomposition,then a five polynomial distribution function is designed to define the discrete value of reference trajectory.In general,three key issues about image features design,multi-variable constraints handling and uncalibrated visual control,are researched for IBVS in this paper.Some feasible ideas and methods are proposed,and this makes important meaning and reference value for developing the theory knowledge of robotics visual control.