• template matching
• dense
• census transform
• augmented reality
• mobile visual computing
• disparity map
• hamming distance
• stereo vision
• real-time
• ARM Mali T600 series
• ARM Mali T604 GPU
• GPU compute
• computer vision
• image processing

Stereo vision is a powerful visual sensing technique to locate, at the proper distance from twin cameras, any specific object in the scene. Nowadays it is not only an attractive subject of study and research in machine vision, robotics and computer vision but it also finds application in many real world case, such as natural user interfaces, industrial automation, autonomous vehicles, and many more. However the stereo vision algorithms are extremely computationally expensive and they result in very high CPU load, particularly if accuracy is needed by the application. In order to get fast reactions from a system, it is really important to ensure data is delivering at high frame rate. Moreover the reliability and robustness of 3D data are important aspects as well to consider especially in a wide variety of scenes and illumination conditions. Many algorithm are proposed every year to improve accuracy/reliability or computation speed: due the high computational complexity, it is still a huge challenge to achieve high accuracy/reliability at high frame rate. A possible solution to overcome this, it is to implement a stereo vision algorithm that exploits the highly-parallel execution feature of a Graphic Processing Units (GPU). The aim of this thesis work is to present a real-time dense stereo vision algorithm based on a multi-resolution strategy and the modified census transform that is tailored to a mobile graphics processor such as the ARM Mali T600 series GPU. The thesis describes the adapted algorithm and the design decisions taken in order to implement an optimally OpenCL code carried out on ARM Mali T604 GPU. The algorithm robustness with respect to changing illumination conditions is also shown. Computing stereo vision on the GPU realizes a significant speed-up and precision improvement, as proved by the performance analysis (in terms of computation speed and accuracy) on widely accepted reference scenes. Particularly are reached comparable results in terms of accuracy with other state-of-art related real-time implementations and approximately 120fps on 320x240 image resolution with 60 disparity levels or 40fps on 640x480 with the same disparity range on above mentioned ARM Mali T604 GPU. This project has been developed with a joint cooperation between the Dept. of Information Engineering of the University of Pisa (Italy) and ARM Holding (Media Processing Division, Cambridge - UK).