Disparity map computation speed comparison for CPU, GPU and FPGA implementations [articol]

Abstract

In this paper a comparison of the processing speed of the disparity map computation using a CPU, a GPU and an FPGA is presented. First the straight-forward implementations of the block matching algorithm for the CPU and GPU are presented, followed by the newly developed architecture for FPGA implementation. The GPU used in this paper is an Nvidia Tesla C1060, programmed using the Nvidia CUDA API. The sum of absolute differences (SAD) has been chosen to compute the matching cost for the block matching algorithm, because of its simplicity, which facilitates a hardware implementation and makes the algorithm suitable for use in applications where a high frame rate is required. The last part of the paper presents a comparison between the processing speeds of the three considered devices.