The Research Of Video Motion Analysis Based On Optical Flow Model

Motion analysis technology is an important part of the image processing. Based on modelof the optical flow, motion analysis technology is investigated in this paper. Optical flow is theproblem to solve the velocity field of a moving object.In general, additional optical flowconstraints are required besides constant intensity assumption. Such as global smoothness,block matching and so on are all optical flow constraints in different aspects. Optical flowmethod is sensitive, highly accurate, and able to represent the velocity field of the object.Motion estimation is based on the method. Optical flow method has been developed as anadvanced motion analysis method.In the analysis part, we investigate both motion detection and optical flow algorithm.Motion detection includes frame difference method and background subtraction. In the study ofthe optical flow method, we introduce optical flow including the gradient method, phasemethod, and block matching method, and then focus on Horn-Schunck algorithm andLucas-Kanade algorithms. Horn-Schunck algorithm uses additional global smoothnessconstraints to solve optical flow problem, and combine the pyramid strategy, and propose amulti-scale improvement of Horn-Schunck algorithm.In the preparation part, we extract each frames of video, and convert video frame into asingle image. The image pre-processing is a main process to eliminate the effect of noise on theresults. Median filter is used in this stage.After studying classical optical flow algorithms, it’s found that the classical Horn-Schunckalgorithm is easily affected by noise; In the flat region, Horn-Schunck algorithm is unable togive an accurate estimation about the optical flow of the image, and unable to give the correctresults when the object has a large-scale displacement. So in this paper we try to integrateHorn-Schunck and Lucas-Kanade from two different directions. A direction is to add localconstraints for Horn-Schunck algorithm, the other is to apply multi-scale levels of image inoptical flow algorithm.Because Lucas-Kanade algorithm is based on local constraints, the optical flow must becomputed after local constraints. The data term and smoothness term in Horn-Schunck equationare calculated separately, and don’t affect each other, so it is necessary that a Gaussianconvolution is added to data term of Horn-Schunck equation. It does not affect the global constraints, but also impose new local constraints. Because the local constraints is added beforecalculating optical flow in the regions, the method in the paper is expected to be insensitive tonoise with a high noise resistance.In the implementation part, we try to apply pyramid strategy in Horn-Schunck algorithm.Horn-Schunck is based on the pixel-wise calculation. With converting the image into differentlevels, we can use the result gained from low resolution as the input of high resolution, and thengradually approach to the final estimation. The experimental results comparing with traditionalHorn-Schunck algorithm shows that the method can produce motion estimation of objects morerobustly and more accurately. The method has a good effect for large-scale motion, and also cansuppress the phenomenon of aperture.At the end, there is a conclusion about the work in this thesis.