Abstract

Visual servoing is a robot control method in which camera sensors are used inside the control loop and visual feedback is introduced into the robot control loop to enhance the robot control performance in accomplishing tasks in unstructured environments. In general, visual servoing can be categorized into image-based visual servoing (IBVS), position-based visual servoing (PBVS), and hybrid approach. To improve the performance and robustness of visual servoing systems, the research on IBVS for robotic positioning and tracking systems mainly focuses on aspects of camera configuration, image features, pose estimation, and depth determination.
In the first part of this research, two novel multiple camera configurations of visual servoing systems are proposed for robotic manufacturing systems for positioning large-scale workpieces. The main advantage of these two multiple camera configurations is that the depths of target objects or target features are constant or can be determined precisely by using computer vision. Hence the accuracy of the interaction matrix is guaranteed, and thus the positioning performances of visual servoing systems can be improved remarkably. The simulation results show that the proposed multiple camera configurations of visual servoing for large-scale manufacturing systems can satisfy the demand of high-precision positioning and assembly in the aerospace industry.
In the second part of this research, two improved image features for planar central symmetrical-shaped objects are proposed based on image moment invariants, which can represent the pose of target objects with respect to camera frame. A visual servoing controller based on the proposed image moment features is designed and thus the control performance of the robotic tracking system is improved compared with the method based on the commonly used image moment features. Experimental results on a 6-DOF robot visual servoing system demonstrate the efficiency of the proposed method.
Lastly, to address the challenge of choosing proper image features for planar objects to get maximal decoupled structure of the interaction matrix, the neural network (NN) is applied as the estimator of target object poses with respect to camera frame based on the image moment invariants. Compared with previous methods, this scheme avoids image interaction matrix singularity and image local minima in IBVS. Furthermore, the analytical form of depth computation is given by using classical geometrical primitives and image moment invariants. A visual servoing controller is designed and the tracking performance is enhanced for robotic tracking systems. Experimental results on a 6-DOF robot system are provided to illustrate the effectiveness of the proposed scheme.