Research On Recognition And Three Dimensional Location Technology Of Underwater Target Based On Monocular Vision

The underwater optical vision technology is one of the important field of the underwater robotic researching. As the basis and key technologies for underwater robot detecting and operating, the automatic target recognition and three dimensional localization techniques have been one of the hot issues of underwater robot research, attracting more and more attention from scholars at home and abroad. The research on underwater optical vision technology shows important research significance and practical value for improving intelligence level of underwater robot.Aim at the target recognition and three dimensional techniques, and on the base of pretreatment of underwater images, this paper carries out some specific research works about shape feature extraction, texture feature extraction, feature fusion, three dimensional localization of structured light and system calibration.In the process of pretreatment of underwater images, aim at the question that the traditional median filter may bring serious distortion to images when the probability of impulse noise is large, this paper introduces adaptive median filter. This method uses the window whose size can change, and judges whether the median pixels in the window and filtering pixels are the impulse noise, then disposes them respectively, which reduces distortion and protects details. The general fuzzy enhancement methods on the base of traditional structure feature and gray feature are inadaptive to underwater images with extremely low contrast and blurry texture. Aim at this problem, this paper brings forward a method to determine the enhancement threshold based on the maximum entropy theory, and carries out fuzzy enhancement in transform domain to increase contrast between background and object. Experimentation validates the pretreatment method brought forward in this paper is effective. In the research on extracting feature of object, on the base of the traditional Hu moment invariants, the new Hu moment invariants are structured newly in this paper. This method considers radial distortion of camera on the base of original moment invariants, so the gray value and position of pixels can be correspondent batter, which improves the clustering ability of eigenvectors to object; When extract texture feature, the Statistical method based on normalized histogram and spectrum method based on Fourier transform are combined, and then a representation method combined with shape feature and texture feature of the object is presented in this paper, which bring the preciser and more complete judgement than single feature. The target cognition classifier is designed in this paper based on improved BP neural network. Methods presented in the above paper are proved effective according to the feature extraction and target cognition experiments.In the researching of three dimensional localization for underwater target, aim at the matching in stereo binocular vision is difficult, a three dimensional localization method based on monocular vision is presented in this paper, which combines line shape laser emitter and underwater camera. The three dimensional coordinates are obtained by parameters calculating while not by on-line estimation, and the underwater refraction infection is corrected, which reduces the distortion of depth information and improves the localization precision. In the research of the key technology to extract center fringe of structured light, considering the infection induced by scattering and refraction that make the intensity doesn't accord with normal distribution, which leading to the exact location of fringe center can't be extracted by the traditional gray weighted centroid algorithm, so a improved extraction method is presented in this paper in which threshold method and direction-variable template technology are combined. This method plays the advantages of both the method fully to improve disposal speed and extraction veracity. The locating experiments of underwater target show that the locating method and improvement work presented in this paper is available to improve three dimensional locating precision of the targets.In the researching of calibrating camera and locating system, aim at the problem that the traditional method brought forward by Ma Songde ignored radial distortion of the camera and the translational motion in three dimensional spaces is difficult to realize, a new method to calibrate camera parameters based on plane translational motion is presented in this paper, and the camera radial distortion factor is considered, which improves the locating precision; For calibrating the structure parameters of structured light system, a calibrating method for monocular vision and structured light locating system is presented in this paper, which takes the translation transformation as the thought and simplifying calibration structure as main purpose to reduce cost. The locating experiments of underwater target show that the system calibration method presented in this paper can improve three dimensional locating precision of the targets availably.