Abstract
The MirosSot and the AndroSot soccer robots have the ability to recognize, and navigate within, their environments without human intervention. An overhead global camera, usually at a fixed position, is used for the robot's vision. Because of the lens distortion, images obtained from the camera do not accurately represent the robot's environment. The distortions affect the coordinates. A technique to calibrate the camera is required to transform the skewed coordinates of the objects in the image to the physical coordinates, which define their real-world position. In this study, a method is proposed for camera calibration using an artificial neural network (ANN) in a two-step process. First, ANN was used to select the camera height and the lens focal lengths for high accuracy. Second, ANN was used to map a coordinate transformation from the camera coordinates to the physical coordinates. During the learning process, the weight of each node in the ANN model changed until the best architecture is reached. The experiments thus resulted in an optimum ANN architecture of 2x4x25x2. The accuracy and efficiency of the camera calibration method were obtained by relearning using the ANN whenever changes to the environmental occurred. Relearning was done using the new input data set for each respective environmental change. Based on our experiments, the average transformation error of the calibration method, using many types of camera, camera positions, camera heights, lens sizes, and focal lengths, was 0.18283 cm.