Estimation Of Optical Flow And Reconstruction Of 3D Structure From Image Sequences

Visual motion analysis is a systematic method on the analysis and reconstruction of 3D motion and structure from image sequences. It utilizes the knowledge of computer vision, artificial intelligence, image process and so on. It is widely used in industrial, military and medical fields. The main ways of visual motion analysis can be classified into feature based and optical flow based.This dissertation concentrates on how to estimate optical flow and reconstruct 3D structure and motion from image sequences.In this dissertation, a new method of optical flow estimation in a color image sequence, based on the Hue and Saturation information, is put forward from the models of illuminated objects. Instead of the straightforward method of using the color components R, G and B as separate images of the same scene, the presented method exploits Hue and Saturation as color invariance under motion. The two color systems HSV and HLS are used for two reasons. Firstly, they have advantages over the traditional method in using the ratio of the color components as separate images of the same scenes so that they can avoid linear correlation to color components. Secondly, they accord with human vision since the two color systems not only have been developed for users but also can represent the color characteristics.This dissertation integrates two ways of visual motion analysis by using the corner points as feature points and estimating optical flow from image sequence. The optical flow is estimated by measuring the displacement of sparse located corner points between consecutive frames. A theoretical proof is given that optical flow can be replaced approximately by using the displacement field. The presented method not only decreases the calculating time in the following reconstructing 3D structure and motion but also solves the problems of tracking and matching image features in the both ways of visual motion analysis.A linear method is derived to determine 3D structure and motion using corner point optical flow. The method has advantages over the traditional methods in algorithm. Some experiments are designed to validate the linear method.This dissertation also describes the implementation of a non-linear algorithm whose uniform observability, minimal realization and stability had been proven analytically by previous works. Experimental results show that the algorithm can provide a good estimation of the 3D motion and structure.Straight-line features are prominent in most man-made environments. They can be detected and tracked easily in image data. Additionally, they provide a great deal of information about the structure of the scene. These properties make them ideal candidates for use in structure from motion algorithms.Some known methods of line segment are integrated and an improved method is given. Canny edge detector is used in this method to get edge, link the edge using Freeman chain code, and then gradually delete useless chain code. Lastly least square method is used to fit for the straight line. The improved method can extract so much information as possible from original images. It excels both in getting the information about end points and in representing the straight line accurately.Based on Hough Transform, a new straight lines tracking method is presented. Rather than tracks straight lines in image space directly, the given method makes use of duality theory in Hough Transform in order to transform line tracking in imagespace into point-tracking in Hough space. The geometrical explanation of the method is provided.The dissertation gives a mathematical definition on straight-line optical flow. By using 21 straight lines optical flow in 3 consecutive image frames, a set of equations about 3D motion parameters were established. After solving the above equations, 12 motion parameters of 3D object and coordinates of the 3D straight line could be obtained. In fact, it is difficult to get the 21 straight lines and their correspondence. A minimum of 6 straight lines optical flow in 3 co...