Binocular Vision-based Moving Target Positioning

Moting-target location based on binocular vision has clear advantage over other methods, which makes it one of the important research parts of computer vision. However, it is a complex inverse process to deduce target position from a pair of 2-D images that are synchronously grabbed from the target at different visual angle. In this thesis, the key technologies such as camera calibration, feature extraction and stereo matching involved in binocular vision were studied.Based on analyzing for the existing methods of camera calibration, an improved plane calibration method was presented by combining conventional calibration method and self-calibration method. The method used linear camera model for calculating initial interior and exterior parameters firstly, then the radial and tangent distortion were introduced, and finally, the nonlinear optimizing method was used to solve distortion coefficients. The method not only overcome the conventional method's faults such as high requirement for facility and complex operation, but also attained higher precision than self-calibration method.Being aimed at the faults that Harris corner detection method needs manual threshold-set, which is prone to cause the corner points' distribution nonuniform or corner point cluster, and could only attain pixel dimension precision, the thesis presents an improved Harris corner detection method. It adopts the measure of image partition and eliminating adjacent corner point to achieve automatic threshold-set, which ensures corner points'uniform distribution; then the Forstner detection method was used to raise the precision to sub-pixel dimension.Stereo matching is the most difficult link within a binocular vision system. As the epipolar line of a practical system is curved, which makes that the searching corresponding points along epipolar line not only is time-consuming, but also is seriously effected by noise. To overcome these problems, the thesis presents a matching algorithm based on image rectification and gray correlation. The curved epipolar line was converted into parallel straight lines, which simplified the solving process of epipolar line, and then the algorithm based on gray correlation was used to match corresponding points. By doing so, the accuracy and stability of stereo matching had been effectively improved.Being directed at the image distortion resulting from the aberrations of actual optical system, a preprocessing method to compensate image distortion effectively is presented. According to the conjugate relation between the object and its image, the object plane was divided into many micro object elements that have one-to-one correspondence with the pixels of image sensor; Monte Carlo method has been used to simulate the light distribution on image plane made by the light from each object element, namely point spread function. A point spread array was constructed from all point spread functions of all elements; each row of the array corresponds to the point spread of an object element. A linear system of equations to describe the relation between all object elements and all image pixels was built on the basis of the point spread array, and a lifelike image was obtained by solving the linear system of equations. The method is particularly useful for the imaging system in which the conjugate distance of object and image is fixed.Finally, the program for binocular vision target location was developed by use of Visual C++. Target location experiments were carried on a self-built experimental system. The results verified the feasibility and the validity of the theoretic approach presented in this thesis.