Research On Detection Algorithms Of The Robot-Oriented Multi-vision System And Its Implementation

To be the newly-developing robot, the parallel robot has many advantages that other robots don't have, such as high stiffness, low cumulative error, simple and reliable structure, large load, fastness and easy control. Nowadays, there has been the system that introducing a single camera into the parallel robot, but the system can only localize and classify the simple 2d shape object. Thus, it is far from the target of flexible manufacture.In this dissertation, the parallel robot is taken as the research object, the multi-vision system is proposed to help the robot sense the 3D environment. It is am at improving the level of automation, the precision of the object detection and the area of application of the parallel robot. How to make the multi-vision system sense the 3D world and how to transform the 3D coordinates to the parallel robot system has much relationship with the corner detection problem, parameter estimation problem and the object detection problem in the complicated environment are deeply analyzed when the multi-vision system is integrated. Based on the detailed analysis of the strength and weakness of the traditional methods, the corresponding solutions of the above problems are proposed in order to lead the parallel robot to detect the object in the complicated environment with high precision. At last, the parallel robot oriented multi-vision detection system is accomplishedFirstly, the kinematic model and integral construction of the parallel robot are explained in detail, the imaging theory of the camera and the operating principle of the multi-vision detection system are introduced. Based on the character of the mechanical structure of the parallel robot, a theoretical model of the multi-vision detection system of the parallel robot is designed.The corner detection is to provide the original data to calculate the internal and external information of the multi-vision system. In order to improve the level of automation and accuracy of the corner extraction algorithm in the process of camera calibration, the problems faced during the calibration when using the circle pattern calibration plate and chessboard pattern calibration plate are analyzed and summarized in detail in this dissertation. With respect to the circle pattern calibration plate, the projection error model of ellipse center is proposed and deduced, each of the influencing factors of the camera calibration are analyzed and summarized by simulation experiments. For chessboard pattern calibration plate, traditional open source calibration codes are analyzed. Unlike the tedious manual operation by using the Matlab calibration toolbox and the failure by using the OpenCV software in the complex environment, an improved symmetry and variance sub-pixel corner detection algorithm is proposed and its feasibility verified by experimental results.The parameters calculation is to estimate the internal and external parameters of the multi-vision detection system. When these parameters are obtained, the 3D coordinates of the object can be estimated and can transform to the parallel robot system. In order to calculate the external parameters of parallel robot oriented multi-vision system with higher precision, a hybrid global optimization algorithm based on the convex relaxation algorithm is proposed in this dissertation according to the character of the rotation estimation problems. The algorithm is applied to solve the single rotation estimation problem for the rotation transformation between the camera and the parallel robot and the conjugate rotation robot for the rotation transformation between each two camera of the multi-vision system. The algorithm combines the merits of the interleaving algorithm and the convex relaxation algorithm. It is proved by the experimental results that this algorithm can converge to the global solutions of the rotation estimation problem fast and accurate.Due to the applied range of the parallel robot such as object localization and classification, the industrial products are usually colorless, textureless and the intensity of the image is easily affected by the site conditions, the only character of the industrial object that can be detected is the shape. In order to increase the ability of detecting different shape of object of the parallel robot in the complex environment, a fast object detection algorithm based on the 2D dimensional shape modeling algorithm is proposed according to information provided by the medial axis or the skeleton of the object by analyzing the traditional object detection algorithm. In order to model various shape automatically, the definition of the shape complexity and the information criterion are introduced in this dissertation. In order to make the definition invariant to rotation, translation and scale, the moment invariance is introduced. Because the image are discrete two-dimensional date, to solve the problem when the scale invariance feature of Hu moment invariance failed for the discrete cases, the improved discrete Hu moment invariance is proposed. To solve the heavy calculation and large costing memory problem, the geometric invariant-shape angle is introduced. The proposed algorithm does the preliminary screening for the outlines of multi-object on the whole, then use the model information provided by the ellipse modeling algorithm to match exactly and detect the object fast and accurate in the case of occlusion, noise and projection finally.In the end, based on the research and proposed algorithm mentioned above, the multi-vision detection system of the parallel robot is developed. The experimental results proved the feasibility and validity of the algorithms proposed in this dissertation.