Histograma de gradientes orientados utilizando processamento paralelo em GPU

Abstract

Feature extraction is fundamental to the process of pattern recognition. In this project a GPU implementation of an image feature descriptor based on local distribution of gradient orientation, the Histograms of Oriented Gradients (HOG) descriptor, and it is compared to pre-existing reference implementations in both CPU and GPU, together with a CPU implementation developed in this project. A speedup of 14,3, compared to the CPU implementation, is achieved for Full HD (1920 × 1080 pixels) resolution images and a performance gain of over 30%, compared to the reference GPU implementation.