The Optical Flow Field Calculation And Its Some Optimization Technologies

Calculating the optical flow field of image sequences plays an essential role in the research ofcomputer vision for motion, which is the important tool of the motion analysis and understanding.Although it lasts for30years, the research has been prompted by the new requirement and the newproblem for the next generation of optical flow field. The new theories, technologies and methodsspring up in this research field, and it has been the hot spots. This thesis aims to improve the qualityof the optical flow field and chooses some frontier and hot problems of this research to breakthrough, which are the large displacement optical flow, the feedback control of estimating procedure,the graph cut optical flow, and detecting and processing occlusion. The multiple disciplines, such asthe optimization, the control, the image processing, and the artificial intelligence, are involved.The major work is outlined as following:(1) The state of the art on the optical flow calculation and the main difficulties and problemsare overviewed. The classification and principles of continuous optical flow estimatingalgorithms, the representation, the test image databases, the ground truth, and the qualityevaluating system are introduced. All above are in favor of the evaluation of the algorithm.(2) On the basis of the technical details about the global and local differential optical flowestimating algorithm, a new local estimating algorithm with optimal controlling ispresented. The controlling procedure of estimating optical flow, the controlling stage, thecontrolling parameters, the objective functions and optimal algorithm are illustrated. Theidea of feedback controlling is introduced into the optical flow calculation, and it isenlightening to improve the precision of the present methods.(3) The mechanism-based error, namely the nonlinearity of optical flow constraint equation, oflarge displacement optical flow calculation based on the HS algorithm is analyzed. Thisthesis illustrates the detail of Harris detector and comes up with a new kind ofcompensated HS algorithm based on matching Harris corner points. The experiment showsthat compensating method is effective.(4) The thesis significantly illustrates the fundamental theory of network flow and graph cut.Three concrete instances of graph cut algorithm, namely the expand move graph cut,the swap move graph cut and the simplified multi-label graph cut, are used tocompute corresponding optical flow field. Above three optical flow fields are comparedwith one acquired by other algorithms in energy drop, motion boundary protection,smoothness. (5) The thesis especially researches the smoothness performances caused by the space term ofgraph cut. Theory and experiment justify that the disturbance of the space term candecrease outlier points. In addition, the discriminating formula for occlusion region isestablished, which is based on the forward and backward optical flow field. A novelnetwork flow graph, which denotes the new space term and energy function, is presentedto control directional smoothness, and so as to deal reasonably with the occlusion.