A real-time stereo matching algorithm using graphic hardware and hierarchical method

This paper proposes a real-time stereo matching algorithm implemented in the graphic hardware. The likelihood model is parallelized and implemented using GPU programming for real-time operation. And the prior energy model is proposed to improve the accuracy of disparity estimation. First, the likelihood matching based on rank transform is implemented in GPU programming. The shared memory handling in graphic hardware is introduced in calculating the matching errors. Once an initial disparity map is determined based on the likelihood model, then the disparity map is iteratively updated by the prior model of disparity field. The prior model reflects the smoothness of disparity map and is implemented by a pixel-wise energy function. The disparity is determined by minimizing the joint energy function which combines the likelihood model with the prior model. These processes are performed in the hierarchical successive approximation approach. The disparity map is interpolated using color-based similarity. This paper evaluates the proposed approach with the Middlebury stereo images. According to the experiments, the proposed method shows good estimation accuracy with more than 30 frame/second for 640×480 images and 60 disparity range. The proposed method is expected real-time stereo camera systems to be popular in the usual PC environments.