Improved optical flow estimation methods using Atomic Functions

The problem of recognizing the objects movement in a sequence of video images has been addressed using different models. Such models attempt to describe the general characteristics of such movements, which require an accurate estimation or the optical flow or velocity of the images. That is the approach of the object movement defined as the projection of 3D surfaces point speeds on a 2D plane. Among several approaches proposed to this end, the optical flow estimation methods from Horn-Schunck and Lucas-Kanade are some of the most widely used. These methods strongly depend on the use of a smoothing Gaussian function. This paper improves the performance of both methods replacing the Gaussian function by a set of functions called Atomic Functions, that present frequency responses with higher attenuation of its side lobes. The performance of conventional and proposed modifications is evaluated comparing the disparity map of the optical flow images obtained, using the Structural Similarity method.