3D Information Extraction Based on GPU

Our project starts from a practical specific application of stereo vision (matching) on a robot arm, which is first building up a vision system for a robot arm to make it obtain the capability of detecting the objects 3D information on a plane. The kernel of the vision system is stereo matching. Stereo matching(correspondence) problem has been studied for a few decades; it is one of the most investigated topics in computer vision. A lot of algorithms have been developed, but only a few can be applied in practice because of the constraint from either accuracy or speed requirement. After the vision system is built, one can get some insights from it, and determine which part of the vision system needs to be improved through experiments. The result shows that the accuracy of current block matching algorithm is enough to be applied in specific environment. Thus, the focus of the afterwards optimization for the currently built vision system is mainly from speed acceleration aspect. After measuring each stage time cost of 3D sensing part of the vision system, the most time consuming stage is from the stereo matching which generates the disparity map or depth map. At last, the stereo matching part is executed on GPU(Graphic Processing Unit) in order to get some performance enhancement, the final result demonstrates that GPU can make the algorithm run in real time, and it is an ideal platform for the further application development of stereo matching algorithm. Because the original speedup of GPU against to CPU is round 35 times at least for desktop GPU, and the optimized speedup of GPU against to CPU can be more than 100 times at least for desktop GPU. 3D Information Extraction Based on GPU