engleski

A Hybrid Local Method for Stereo Correspondence based on Three-Dimensional Recursive Search

Obtaining dense disparity maps of stereo image pairs is a common problem in computer vision. We present a new hybrid method for dense stereo correspondence based on the well-known winner- take-all optimization and the three-dimensional recursive search (3DRS) algorithm. Originally devised for motion estimation and frame rate conversion in high definition televisions, the 3DRS algorithm employs predictive block matching to generate a coarse correspondence map between two images in a sequence. The algorithm is widely used in ASIC and FPGA implementations, achieving real-time performance in processing of 60Hz high definition video signals. In our proposed hybrid method, the 3DRS algorithm is applied to a stereo pair in order to generate a coarse disparity map, which is further optimized using a local, winner- take-all driven coarse-to-fine approach. The selected approach greatly reduces the local disparity space to be searched, yielding significantly better run-time performance while retaining or even improving the accuracy of the results. The method is evaluated using the Middlebury stereo datasets and compared with both the reference winner- take all algorithm and coarse-to-fine hierarchical optimizations. The obtained results favor the proposed method in terms of accuracy and performance. The proposed method is suitable for implementation in custom parallel computing hardware, ASIC or FPGA.

3D reconstruction ; passive stereo ; active stereo ; winner take all ; 3DRS ; three-dimensional recursive search

nije evidentirano