Accurate motion estimation based on moment invariants and high order statistics for frames interpolation in stereo vision

Abstract

At present, stereo video sequences are actively used in the movie industry, geographical information systems, and in navigation systems, among others. A novel motion estimation method simplifies the frames interpolation by forming an accurate set of local motion vectors in neighbour frames. First, the motion in a scene is estimated by block-matching algorithm roughly. Second, the accurate estimations are calculated according to type and velocity of motion. The Hu moments are used for a fast and transition motion in a scene. The Zernike moments are applied for fast and noisy scenes with a transition/rotation motion. The high statistical moments (kurtosis particularly) are reasonable for the accurate analysis of a slow motion. Such approach provides a smooth motion into additional interpolated frames that improves significantly the final stereo video sequence. Experimental results show the efficiency of the proposed method for frames interpolation. The detection of local motion vectors achieves 86% accuracy by usage the Zernike moments and 88% accuracy by a kurtosis calculation.