A neural network approach to camera calibration for movement tracking

Abstract

In this paper, neural networks were trained to obtain the transform between pixel coordinates and real world coordinates. We determined that it is possible to calibrate using neural networks utilizing a combination of points that both lie near the center and around the edges of an area of interest. A minimal number of 33 input and output pairs can be used to successfully calibrate a playing field with an average Euclidean distance error of 1.28 cm. We demonstrated our technique by calibrating a Foosball table using 23 points. We tracked a ball moving around the table and then using the transform learned by the network, we were able to obtain the actual coordinates of the trajectory.