Research On Optical Flow And Scene Flow Estimation

Optical flow is the surface movement of a picture’s gradation pattern.It is the projection on a two-dimensional plane of a three-dimensional image. While scene flow is a three-dimensional movement field formed from moving scenes in space. Both of them have a wide range of application. Robustness and accuracy are the key issues in the existing optical flow estimation.A novel method based on texture is proposed here to improve the robustness. Meanwhile, two methods of scene flow estimation are introduced, which overcome the existing problem in scene flow estimation,including descending the errors caused by camera motion. The main research results obtained are as follows:First, classical global estimation of optical flow, Horn-Schunk method, is introduced in detail. A basic model naming “Horn-Schunk- multi- scale- median filter” is built using multi-scaled strategy and median filter, which can over-come the difficulty of large displacement in the estimation of optical flow as well as assure the accuracy.Texture decomposition is a useful way in image processing. The paper uses texture of images in the computation instead of original images. Experiment results show that it cannot improve the computational accuracy. However, it is better in dealing with the problems of noise, illumination variation, occlusion,.etc. Compared with classical Horn- Schunk method, the new model is more robust.Secondly, based on the basic model ”Horn-Schunk- mult i- scale- median filter”, the algorithms are extended to scene flow. O ne method is using left and right photos taken by the binocular stereo vis ion system. Another way is using RGB images and their depth maps. Both of the two can compute the scene flow.Finally, to deal with the translation of the camera, a new algorithm is designed,which can efficiently distinguish the global motion and local motion.