Accelerating Stereo Matching on Mutlicore ARM Platform

Abstract

Stereo vision is a well-known technique in computervision used to acquire the 3D depth information of a scenefrom two or more 2D images. One of the main issues with anystereo vision system is how to make a good trade off betweenthe processing speed and the quality of the disparity map. Thisissue can be resolved through the use of dedicated hardwareplatforms, like Field Programmable Gate Arrays and GraphicalProcessing Units, which are considered as expensive solutions.In this work, the challenge of accelerating stereo matching onlow cost multicore platforms is tackled. We present a novelsoftware implementation of a sparse Rank algorithm, that uses amodified Sum of Absolute Differences 1D box filtering algorithmin the correlation stage. Consequently, we reduce the numberof computations and memory space needed for computing thedisparity map. The system is implemented on a multicoreAdvanced Risc Machine platform (ODROID XU4). Experimentalresults show that the system is capable of achieveing a processingspeed of 59 Frames Per Second for images of size320×240pixelswith a disparity range of 20 pixels. Furthermore, the sparse Rankstructure does not affect significantly the overall quality of thedisparity map.