Abstract:For an agricultural robot, camera vision implements the function of detecting ripeness, shape, size of crops and locating the stems precisely in order to harvest with no damage. A common calibration model with adjustable accuracy was needed to achieve higher precision in cooperative works of certain manipulator and CCD cameras. Creatively, a variable order NURBS surface model for camera calibration was presented. Firstly, seven images were acquired by a slide way and a calibration board of A0 size. The data was extracted afterward. Based on the thought of NPBS method, four double NURBS surface calibration models with order of 3, 4, 5 and 6 were setup respectively using data of five images (index:1, 3, 4, 5 and 7). Other two images (index: 2 and 6) were used to evaluate the calibration error. According to the requiring preciseness and the calculating time, the models of order 3 and 4 were chosen to accomplish model switching (higher order models can be chosen according to higher precision requirement). Secondly, camera image plane was subdivided evenly with respect to parameter (u, v). In each subdivision, an arithmetic average deviation of calibration was calculated by using the model with lower order. Then a threshold was determined, and the diagram of model switching was formed. That meant using a higher order NURBS surface model in the higher distortion region, vice versa. Thus, orders of the NURBS surface model can be chosen according to subdivision of the image plane based on preevaluation of the calibration error. In experiments，it was proved that the average calibration error was under 0.89mm. It was accurate enough for our agricultural robot prototype.
