Fiducial Marker Detection Using FPGAs

This work presents an approach for computing camera pose from images of fiducial markers. The processing is shared between an FPGA and a PC: the FPGA converts grayscale camera images into binary format and transfers them to the PC; and the PC detects fiducial markers in the received binary camera images and computes the position and orientation of the camera. The approach is evaluated in terms of camera pose error and processing time by using a 3D simulator and subsequently tested in our Immersion Square environment. The 3D simulator has been developed in course of this work in order to enable a thorough theoretical evaluation. It is used to generate a large number of test images from different camera positions by simulating a subset of the Immersion Square. Synthetic images generated in this way undergo a set of transformations in order to make the evaluation closer to images obtained from a real camera by using Gaussian blur of different kernel sizes, and several types of noise. The results of theoretical evaluation show that depending on the chosen parameters, a relatively high frame rate of approximately 60 fps can be achieved. Moreover, the parameters allow for a trade-off between speed, accuracy, and precision. The insights gained from the theoretical evaluation are used to develop an optimized system that has been subsequently tested in the real Immersion Square. The tests show that the optimized system runs with the average of approximately 80 fps and is capable of achieving higher frame rates of 100 fps. It displays better accuracy, and higher precision than the unoptimized version. In addition, the real world tests show that the system produces very stable crosshairs, such that it can be used for interacting with the Immersion Square in applications where precise pointing is required.