Approach To Image Sequences Analysis Based On Sparse Optical Flow Field

The analysis of dynamic image sequences is a quite hot topic in computer vision, and its basic task is to estimate 3-D motion and structure parameters from a sequence of images. There are two basic analysis approaches: the feature-based approach and the optical flow based approach. Because the classical optical flow algorithm is sensitive to noise and time-consuming, it is not extremely suitable for tracking timely. Thus, we adopt sparse optical flow approach by combing that two approaches in this paper.Firstly, in this paper, we develop video capture program completely by software method with VFW software package provided by Microsoft Corporation, and in this way the generality of the program is improved.Secondly, in order to eliminate image noise, separate the object from background and extract edge points of the object, a series of pre-processings are carried out to the images captured, such as image smoothing, threshold segmenting and edge tracking. This makes preparations for the computation of sparse optical flow field.Thirdly, we extract feature points of the object quickly and exactly by the method of combining area constraint and angle constraint, and use the Hopfield neural network with massively parallel processing capability to perform relaxation matching of feature points within sequences of images. On this basis we compute sparse optical flow of the feature points by means of parallax.Finally, we estimate 3-D motion and structure parameters of the body by sparse optical flow. In order to solve the monocular vision problem of estimating depth estimation, we estimate the depth of the object by means of position-from-defocus, and carry out depth adjusting to meet the measure precision demand. Under the condition of acquiring depth information, 3-D motion parameters of the object can be estimated uniquely only using 5 feature points with a robust algorithm of recovering 3-D motion parameters of the object from optical flow field.