Non-local weighted regularization for optical flow estimation

Global optical flow characterizes the overall motion while local optical flow characterizes the individual's movement. Therefore, both global and local optical flow is important to analyze the motion of each object in an image and of the camera. In order to robustly and accurately estimate the optical flow which includes both global and local motion information, this paper proposes a novel variational framework with non-local weighted regularization for optical flow estimation (NLWOF). The proposed NLWOF strategy includes the following two key parts: Firstly, non-local prior is used a regularization to robustly estimate the local optical flow due to that the similar structure patches using in non-local weight can suppress noise interference. Secondly, the solution of the NLWOF model can be approximated by the employment of the Lagrange-Euler formula and the approximation of the Laplace operator. Experimental results in both quantitation and qualitation verify the effectiveness of the NLWOF scheme, and even better than those of the state-of-the-arts.