Abstract
Motion estimation has been increasingly important in the field of computer vision in recent years. The motion estimation is conducted using optical flow field. The estimation model of optical flow consists of data term and the regularization term. In this study, we aimed to present and compare six variational models by varying the composition of the data term. The variation is formulated through application of different penalty functions namely, quadratic, Charbonnier and Lorentzian in data term. The quadratic penalty term has linear characteristic which eases its implementation in the model. The Charbonnier function is deployed to increase the robustness of the model against the outliers, and the motive of utilizing Lorentzian function is to test the behaviour of variational model when nonconvexity is incorporated. Further, these penalty functions are convolved with the non-local weighted function to incorporate global motion information. Thus, these established models are solved using various numerical techniques. In the regularization term of all the presented models, fractional order-based derivative called Marchaud is considered. The Marchaud fractional derivative is discretized using Grunwald-Letnikov derivative method. The result obtained from the solution of each model is visualised using a color map and vector field. Moreover, the variational models are also evaluated and compared using numerous evaluation metrics such as, AAE, AEE, AENG, RMS, and WE. A thorough analysis is conducted on variety of datasets including synthetic and real dataset. The comparison results are employed by the different bar graphs, and statistical analysis is also conducted to ensure the scientific validity.

This publication has 35 references indexed in Scilit: