A Hardware Friendly Stereo Match Refinement Algorithm Using Disparity Gradient Based Region Growth Method

Stereo Match is one of the key fields in computer vision. Although many dense two-frame stereo algorithms have been developed in this domain, few utilize cross check and disparity gradient based refinement method. This paper proposes: (1) Cross check method using two generated disparity maps based on left and right original images. (2) A novel occluded and low-texture region growth method based on disparity gradient. (3) Disparity voting method to reduce random error. These practical methods are hardware friendly and can notably improve match accuracy as well as lower the computational and space complexity. The proposed algorithm reaches the highest match accuracy for HR images among existing local methods, attesting to its outstanding effectiveness.