Anisotropic Diffusion-based Texture Preserving Optical Flow Estimation using an Incremental Multi-resolution Approach

Abstract

The purpose of this study was to propose a texture preserving optical flow approach based on non-linear anisotropic diffusion filtering scheme to estimate the accurate motion fields. The proposed technique estimated the motion fields based on several fundamental techniques. Firstly, we recomposed the original image by using the structure-texture decomposition to remove the structure information from the input image. We then employed an anisotropic diffusion-based slightly non-convex TV-Li minimization scheme into the spline interpolation based coarse-to-fine warping approach. The intermediate bilateral filter is applied after each iterative warping step to prevent oversmoothing effects. We selected the average angular error and the average end point error as evaluation indices to evaluate the improvements of our approach. The results of the present study found that the proposed approach outperformed all employed methods in terms of both AAE and AEPE. Moreover, our approach successfully handled the inherent shadow and shading artifacts. We also obtained the reliable motion fields on the image boundaries and poorly textured regions.