The Study Of Actuator Autonomous Positioning And Precise Control Based On Visual Servo

In the application scenario of safe production emergency disposal, the disposal platform reaches to the operational scope of the actuator under the human intervention. However, the actuator's over-operation,under-operation or the misuse in visual blind zone are caused by the delay,simultaneously, the shape of the target object is different, and the relative position between the target and end-effector is changing. These factors have a significant impact on the actuator's precise positioning and fast compensation of positioning error. The system cost is higher. In order to solve the above problems, these research based on a simple depth imaging equipment are the identification of the target object and its position determination, depth information extraction, the moving end-effector state tracking, positioning error correction between the target and end-effector respectively. The coordinate frame relation is established among the vision-base-end-effector-target. Then, a non-standardized five-revolute open-chain linkage rigid actuator (NSA) is driven to reach the destination. The following research methods are proposed. (1)Kinematics model of NSA was established to generate space motion trajectories of each joint and end-effector. The actuator dynamics model was established to obtain driving force of each joint. The range of the joint actual rotation angle and inverse kinematics expression are determined by "vector-product" method. A new scaling learning vector quantization neural network (LVQNN) was proposed to evaluate space trajectory of the actuator. According to the evaluation result and driving force of each joint, the optimal space trajectory planning was obtained.(2)It is always difficult to establish Jacobian matrix and determine the coordinate-frames of links. This leads to the localization capability analysis difficult. A new analytical method to establish the Jacobian matrix and determine the coordinate frames for joints and links are proposed. The proposed method made the positioning analysis of end-effector easier in space. At the same time, it is necessary to prove the effectiveness of the proposed method theoretically and verify the localization and configuration capabilities through simulation. Inverse kinematics is simplified by artificial neural network based on back-propagation multi-layer perceptron (MLP). The training set used are Denavit Hartenberg (D-H) parameters in Cartesian coordinate frame. The weights are updated continuously which reduces the mean square error(MSE) gradually. The localization function is defined to evaluate the positioning property of end-effector. At the same time, in task space, it will check whether the actuator has reached the target point along the direction needed or not. (3) The world coordinate frame, the base coordinate frame, the joint coordinate frame, the end-effector coordinate frame and the visual coordinate frame are established. And the transformation matrices among the coordinate frames are calculated.Theoretical model of the transformation between vision and end-effector coordinate frame is an optimized equation including multivariate polynomial. The polynomials are optimized globally using convex linear matrix inequalities (LMI). 2D information measured by Kinect is mapped into the base coordinate frame of the actuator. Thus, the position of end-effector can be obtained in real time. At the same time, it can guide NSA for accurate operation. It is necessary for a robot vision system to complete tasks such as precision assembly. (4)The online analysis of target object recognition and end-effector positioning has always been a difficult challenge in the process of actuator autonomous positioning. In order to solve this problem, a simple depth imaging equipment (Kinect)was used as a visual servo system. Two kinds of target recognition algorithms were proposed. One is classification algorithm based on neural network of error back-propagation(BP). The other is image matching based on speed up robust features (SURF). BP tends to identify different classes. SURF tends to identify the objects within the class. The target center (TC) was calculated according to the depth images. (5)During positioning, Kalman filtering method based on three-frame subtraction is proposed to capture the end-effector motion and estimate its dynamic positions. Batch point cloud model was proposed in accordance to depth video stream to calculate the space coordinates of the end-effector. And the end-effector center (EEC) was also calculated. If EEC approaches to TC gradually within an allowable error range, it shows positioning successful. The target point cloud was fitted based on radial basis function (RBF) and morphology to verify the effectiveness of positioning algorithm.The following results can be obtained through simulation and experiments. (1)The space trajactories of each joint and end-effector were evaluated by improved LVQNN with scale transformation. It can guide the actual operation of the NSA. It can provide a reference basis for trajectory planning and precision control of NSA. It indicates that the manipulator can meet the requirement of dynamics by calculating the driving force of each joint and the moment of inertia of each link. (2)The inverse kinematics model is simplified based on MLP. On the premise of meeting the accuracy requirement, the solving efficiency of inverse kinematics is improved. Locatable function can be analyzed quantitatively. The sufficient and necessary conditions for the positioning are proved. And the end-effector's motion range is obtained. When the end-effector was positioned at a point p=(41.4,89.0,104.5)/cm, the minimum value of localization function is reachability(DH)=0.96. The maximum value of cost function is configuration(DH)=4.039e14. (3)The transformation matrix is obtained under the ways of "eye-in-hand" and"eye-to-hand",that is the pose relationship among the base, end-effector,vision and target. The internal and external parameters of Kinect are obtained through calibration (also called the transformation between vision coordinate frame and image coordinate frame). Spatial coordinates of target area are calculated by 2D-3D mapping model. (4)The target objects to be recognized are cylindrical, square and spherical objects.Classification values 0.1, 0.2 and 1 in BP denote the cylindrical, square and spherical objects respectively. The classification results of BP network are one to one correspondence with the actual classification basically. Recognition rate of target objects is 0.99. A cylindrical object is selected randomly from test sample. Its coordinates of TC are obtained.Intraclass identification based on SURF is carried out at two scenarios.Target to be identified is called the test sample. Target that has been registered is called registration sample. The similarity between the test sample and registration sample is described by matching ratio to measure recognition performance. Rotation and scaling transformations are added in the test sample. And the matching ratio, computation time and the number of matching pairs are obtained. (5)The experiments of NSA autonomous positioning and control have demonstrated that the and-effector positioning error can be corrected in a short time. The gradual convergence of the end-effector center (EEC) to the target center(TC) shows that the autonomous positioning is successful. Hence, the proposed algorithm is competent for autonomous positioning, error measurement and compensation. The computational ability is increased and system efficiency is greatly improved.We can get the following conclusions through above research methods and experiment results. NSA can meet the kinematic and dynamic requirements. It has a practical value. High recognition rate shows that the extracted features are comprehensive and critical. It also shows that the algorithm is rational. Autonomous positioning algorithm based on simple depth imaging apparatus is competent for target object recognition, the spatial position determination, able to detect and track the movement of end-effector. And its state estimation is finished. EEC is also obtained in real-time. According to the EEC state of convergence to TC, if EEC coincides with TC within a certain error, it indicates the positioning successful. Otherwise, the motion of NSA will be corrected until successful positioning according to positioning error. Meanwhile,transformation of frames makes end-effector-vision coordinated and coherent. Point cloud mapping of 2D-3D is obtained. Finally, 3D surface fitting is used to verify the effectiveness of the algorithm in the process of positioning.