"Parallel CUDA implementation of a stereo matching algorithm"
Fursov V.A., Goshin E.V., and Kotov A.P.

Searching for the corresponding fragments and points on several images of the same scene is one of the central problems in many applications: autonomous navigation extended landmarks, pattern recognition, 3D-scene reconstruction, etc. To solve this problem, various correlation methods for the similarity analysis of fragments are used. Algorithms based on these methods have a high computational complexity. In this paper we consider a stereo matching algorithm for 3D-scene reconstruction. We propose a computational scheme that improves the performance of this algorithm. This computational scheme is implemented using CUDA technology. A high degree of parallelism is achieved due to a large number of the same operations for corresponding points on epipolar lines. Numerical experiments were carried out using the proposed parallel algorithm. The resulting speed-up is estimated.

Keywords: stereo matching, 3D reconstruction, projective geometry, epipolar geometry, parallel computing, graphics processors, CUDA technology.

  • Fursov V.A. – Korolev Samara State Aerospace University, Faculty of Basic Training and Fundamental Sciences; shosse Moskovskoe 34, Samara, 443086, Russia; Professor, Head of Department, e-mail: fursov@ssau.ru
  • Goshin E.V. – Image Processing Systems Institute, Russian Academy of Sciences; ulitsa Molodogvardeiskaya 151, Samara, 443001, Russia; Graduate Student, e-mail: goshine@yandex.ru
  • Kotov A.P. – Korolev Samara State Aerospace University, Faculty of Basic Training and Fundamental Sciences; shosse Moskovskoe 34, Samara, 443086, Russia; Engineer, e-mail: antonykotov@gmail.com