Anisotropic multi-scale Lucas-Kanade pyramid

Abstract

The Lucas-Kanade (LK) algorithm provides a smart iterative parameter-update rule for efficient image alignment, and it has become one of the most widely used techniques in computer vision. Applications range from optical flow and tracking to layered motion, mosaic construction, and face coding. In this paper, we propose a novel Anisotropic Multi-Scale Lucas-Kanade Pyramid (AMSLKP) method. By extracting image pyramids from the original images and iteratively implementing LK algorithm at each level, the Lucas-Kanade Pyramid (LKP) gained better robustness and accuracy. Moreover, instead of calculating gradients in single direction with fixed scale sizes, this paper introduces anisotropic local polynomial approximation (LPA) and intersection of conference intervals (ICI) method to the LKP. The proposed AMSLKP method first calculates the directional estimates and gradients with multiple scales; then for each direction, it adaptively selects the optimum scale for each pixel in the image using ICI rule; at last, the estimate and gradients of the distorted image are computed by fusing the directional results together. The proposed method is evaluated in different noise conditions with various distortion levels. Experimental results show that the AMSLKP method improves the accuracy by more than forty percent compared to LKP method.