Pipelined FPGA Implementation of a Wave-Front-Fetch Graph Cut System

The current mainstream method for stereo vision is to find corresponding points of two-dimensional images obtained from multiple cameras and restore three-dimensional information using the principle of triangulation. However, the occlusion problem often makes it difficult to search for corresponding points. Therefore, a new approach has been proposed in which the three-dimensional space is directly considered as a three-dimensional graph instead of searching for corresponding points of two images. In this approach, a 3D grid graph is constructed based on luminance values obtained from the left and right cameras, and a highly likely object surface is obtained by cutting this graph. This paper proposes a pipelined architecture for 3D grid graph cut, aiming at a real-time stereo vision system. The system uses Wave-Front-Fetch algorithm, which is oriented for parallel processing. We achieved processing time of about 21 ms for a graph of \(129 \times 129 \times 16\) nodes, resulting in a frame rate of about 49 fps. Our approach was about 19 times faster than a well-known graph cut software library.