Robust Optical Flow Estimation via Edge‐preserving Filter

Although the robustness of optical flow field has been improved over the past few years, the problems of blurring motion boundaries and ill-defined edges have remained unsolved challenges in optical flow estimation. Variational optical flow estimation methods may not handle the abovementioned problems since the local filtering-based optimization may cause the problem of blurring motion boundaries. In this chapter, to address the issues of blurring motion boundaries and improve the robustness of optical flow, we propose a new nonlocal total variation (NLTV)- L 1 optical flow estimation method based on a weighted guided filter (named NLTV- L 1 +WGF). The WGF is based on a linear model, which is computationally efficient and edge-preserving under complex outdoor scenarios (i.e., haze, fog, etc.). The NLTV- L 1 optical flow estimation method is refined by employing the WGF in a nonlocal term and then optimized using the coarse-to-fine strategy. We evaluate the performance of the proposed NLTV- L 1 +WGF optical flow estimation method on two popular optical flow datasets (i.e., the Middlebury dataset and the MPI-Sintel dataset) and on several other outdoor complex datasets (i.e., the Foggy Zurich dataset and the Traffic dataset). Moreover, we evaluate the proposed method in a more complex scenario for the task of ant behavior study. It is challenging to perceive the motion and interactions of ants because the size of ants is small, they usually move quickly, and they may not move for a long period. Experimental results show that the proposed NLTV- L 1 +WGF method can preserve the edges of moving objects and improve the accuracy of optical flow estimation compared with several other state-of-the-art optical flow estimation methods.