Parallelization and Optimization of Feature Detection Algorithms on Embedded GPU

In this paper, we parallelize and optimize the popular feature detection algorithms, i.e. SIFT and SURF, on the latest embedded GPU. Using conventional OpenGL shading language and recently developed OpenCL as the GPGPU software platforms, we compare the implementation efficiency and speed performance between each other as well as between GPU and CPU. Experimental result shows that implementation on OpenCL is more efficient but has comparable performance with OpenGL. Compared with the performance on the embedded CPU in the same application processor, the embedded GPU runs 4⇠5 times faster. Furthermore, we measure and compare the power consumption on each implementation, which shows that OpenCL consumes less energy than OpenGL.