Extension of the BRIEF descriptor for color images and its evaluation in robotic applications.

Abstract

During the development of the project, new extensions of the BRIEF descriptor are defined based in the use of color information in different color spaces. They are evaluated against each other and the original BRIEF descriptor (that uses intensity pixel values in grayscale images) by means of a matching test of different points of two different images of the same scene. The selected extension is used to recognize objects in a real time robotic application by means of a classification method. Therefore, the target is to improve the BRIEF descriptor to be applicable in color images by defining the different color extensions and making a comparative evaluation to ensure selecting the best one that most improve the basic BRIEF descriptor with the use of color information. The speed and the recognition rate are computed for every extension and the basic BRIEF to compare the performance of each descriptor. Ones an extension is selected, it is used the Bag Of Features model to create an algorithm able to recognize objects. The BOF model detects the keypoints with the FAST algorithm, describes them by means of the selected extension of the BRIEF descriptor and classifies them using a linear Support Vector Machine. Finally, the algorithm is tested in different situations and conditions, varying illumination, a little bit rotation and background, and checking performance with occluded objects