Camera Geometric Calibration Using Dynamic Single-Pixel Illumination With Deep Learning Networks

Abstract

Traditional methods of geometric camera calibration (GCC) are based on angle measurements (AM) or diffractive optical elements (DOE). However, the AM-based approach has the low accuracy and reliability because the vibration of rotating mechanisms, i.e., turntables and angle measuring accuracy easily influence the calibration accuracy. The DOE failure easily occurs in the calibration process when some micro-apertures are blocked by dust particles. In this paper, a new method for the GCC by means of single pixel illumination in a deep neural network (DNN) is presented. A closed-loop calibration link is composed of a single stimulus input produced by a single pixel generator and a collimator, an uncalibrated camera, and a DNN. The dynamic single-pixel illumination forms the different stimulus input of the DNN at different stimulus time. The synaptic weights (the camera interior parameters) of multilayer perceptron are adjusted until the cost function is minimized. This method is able to avoid the aforementioned shortcoming of conventional calibration methods. This method can be especially used for on-ground calibration of remote sensing cameras but in principle also suitable for on-orbit GCC and other cameras.