Optical Flow Estimation Based On Models Of Motion Boundaries

Optical flow estimation is one of the basic problems in computer vision and hasreceived much attention. The motion boundaries violate the smooth assumption, whichhave a great impact on the accuracy of optical flow estimation. With the study of motionboundaries problems in the optical flow estimation, scholars have proposed a lot of al-gorithms to improve the accuracy of optical flow estimation. This thesis focuses on theanalyzingandmodelingthemotionboundariesproblem,andprovidesalgorithmonsparsemodeling of motion boundaries and enhancing its sparsity, and algorithm implementationon explicit modeling of motion boundaries based on image segmentation.The global smooth assumption violates because there are different movements ofobjects in the motion scene. The motion of different points on the same object can berepresented by one motion model and the motion parameters are consistent. This consis-tency result that the optical flow gradients are mostly zeros in the same motion region,whilearelargernon-zeros atmotionboundariesbetweendifferentobjects. Thisalsoleadsto the sparsity of motion boundaries in optical flow field. According to the CompressiveSensing theory, the sparsity of signal affects the accuracy of signal reconstruction. Toimprove the reconstruction accuracy of sparse optical flow signal, we propose a opticalflow estimation model with enhanced sparsity, which is inspired by statistics distributionof optical flow gradients and maximum likelihood estimation, and can be simplified toa iteratively reweighted optimization problem. To enhance the sharpness of motionboundaries and robustness of brightness constancy assumption, we further employ a ro-bust function as constraint of reweighted model to obtain the reweighted TV-model.Experiments show that, the reweight scheme from prior optical flow statistics distribu-tion can improve the accuracy of optical flow estimation, in both maximum likelihoodestimation and robust estimation.From another perspective, we consider explicitly modeling the motion boundaries.We divide the field into different motion regions by establishing the Markov random fieldmodel and doing image segmentation. In each region the smooth assumption can be satis-fied while between them the motion boundaries can be obtained explicitly. The occlusionwhich violates brightness constancy assumption can be detected by combining the motionboundaries with optical flow information, and optical flow of occlusion can be estimated by the smooth assumption. This layered model of optical flow estimation not only im-proves the estimation accuracy of motion boundaries and the overall accuracy of opticalflow field, but also detects the occlusions to improve the estimation accuracy of opticalflow in occluded regions.In this thesis, on the basis of analyzing the influence of motion boundaries on theoptical flow estimation and sparsity of motion boundaries, we propose an enhancing opti-cal flow sparsity model with prior statistics distribution and its optimization method, andimplement layered optical flow estimation model based on image segmentation and ex-plicitlymodelingmotionboundaries. Theexperimentresultsprovethatthemoreaccuratemodel for motion boundaries results in the more accurate estimation of optical flow.