Research On Vision Based Configuration Detection For Continuum Robot

With the development of robot technology, research on continuum robots is becoming a hot spot. Feedback control with pose or shape information is a necessary condition to achieve real-time and precise movement for continuum robot. Because there are not joints in continuum robot, conventional sensors for example encoders and potentiometers cannot be used to get its position and attitude. To solve above problem, we build binocular stereo vision system and use SOM neural network to achieve three-dimensional shape detection for continuum robot.Firstly, we establish binocular stereo vision system for continuum robots. Zhang’s calibration method is used to achieve a monocular camera calibration to obtain more accurate intrinsic parameters and distortion parameters of camera. According to the principle of stereo vision, we complete stereo calibration for binocular stereoscopic vision and the relative position parameters of two cameras are obtained, including rotation matrix and translation vector which right camera is relative to left camera.Secondly, this paper proposes plane SOM(Self-organizing Map) algorithm and stereo SOM algorithm to achieve three-dimensional shape detection for continuum robot based on conventional SOM algorithm. The difference between these two algorithms is that the two procedures of cluster analysis and three-dimensional reconstruction are interactive in stereo SOM. Yet three-dimensional reconstruction is only executed once after completing clustering in plane SOM. In order to evaluate these algorithms, this paper proposes quality error and topological error for SOM algorithm. In this paper, the principal component analysis is used to reduce the dimension of sample to improve running efficiency of SOM algorithm. To reflect the ability which three-dimensional shape detected by vision track actual shape, we define average position error.Finally, we detect shape for continuum robot. The best design parameters are obtained through SOM network design experiments, and we use these best parameters to make plane SOM algorithm experiment and stereo SOM algorithm experiment respectively. In contrast, plane SOM algorithm has higher accuracy and is able to get better result when it reconstructs the continuum robot shape. Hence these experiments verify the superiority of the plane SOM. On the other hand, the paper verifies the accuracy of shape detection in plane SOM through maker identification. The experiment result is that the average position error can reach 21.24 mm and the relative error is 2.38%. And we prove that this error derived from non-alignment error.