Main | Publications
My defense happened in August 17, 2012. I was supervised by professor Anna Helena Costa Reali, the LTI laboratory from USP. You can download my thesis, and there is also a couple of videos of the defense on YouTube: Video 1 Video 2, but it is all in Portuguese.
Estimação de orientação de câmera em ambientes antrópicos a partir de edgels
(en: Camera orientation estimation in anthropic environments from edgels)
This thesis presents Corisco, a method to estimate the orientation of a camera from a single image captured from an anthropic environment. Corisco was developed with the objective of answering the needs of Mobile Robotics applications, and of the analysis of large set of images, what means that the method should present not only a good computational performance, but it should also be able to use different camera models, allow to control the compromise between calculation speed and result precision, and must also be capable of both exploiting initial estimates of the result, and of operating without any initial estimates. Corisco presents all of these characteristics. The considered environments have a natural reference system with three orthogonal axes, and contain sets of lines parallel to these axes. The estimated orientation is a three-dimensional rotation between the natural reference frame and the camera frame. Corisco requires the knowledge of the camera model, but any camera model can be used. Corisco analyzes images using a process that extracts edgels, which are points located on the projections of the environment lines, associated with the tangential direction of the line projection at that point. This edgel extraction technique uses a grid mask that can sub-sample the data, creating a compromise between speed and precision. The orientation is estimated through a two-step optimization process that minimizes an objective function defined by the M-estimation technique, using a redescending error function. This technique is equivalent to the application of the MAP or the EM estimation in similar existing methods. The first optimization step uses the RANSAC algorithm, allowing Corisco to work without initial estimates, and the second step is a continuous and constrained optimization process that explores the orientation parametrization by quaternions. Corisco was tested with different camera models, including the perspective projection, a model with radial distortion, and two omnidirectional projections, the polar equidistant and the equirectangular. The mean calculation time can be controlled through a couple of parameters, which may also affect the accuracy. The accuracy observed by comparing the Corisco estimates with reference orientations was typically near 1 degree for execution times above 20 seconds, and approximately 4 degrees for less than two seconds. This performance attained the established objectives, and the experimental results validated the method for practical applications.
Keywords: Artificial intelligence. Computer vision. Pattern recognition. Image processing.