Research On Underwater Target Image Processing And Location Methods Based On Binocular Vision

Autonomous Underwater Vehicle-Manipulator System(AUVMS)is an equipment without pilot and cables to accomplish autonomous intervention in ocean environment.The precondition for AUVMS to accomplish autonomous intervention is to obtain the target information quickly and accurately.Sonar vision and light-based vision are two major ways for acquiring the target information.Compare with the sonar vision,light-based vision have higher resolutions during underwater close-range intervention and is the main way for AUVMS to obtain information.In this paper,on the basis of light-based vision,the underwater target image processing and location methods are researched and have important theoretical and practical value for improving the ability of AUVMS to accomplish underwater autonomous intervention.The research background of this paper-is AUVMS to accomplish autonomous intervention,so the underwater target image processing and location methods based on binocular vision are researched,including image noise reduction,image segmentation and target location.Based on the research of these methods,an underwater target location system based on non-parallel binocular vision is developed.This system is aiming at obtaining target information accurately,stability and quickly for cooperating the AUVMS to accomplish real underwater autonomous intervention(briefly referred as the intervention in the following).The contents of this paper are summarized as follows:(1)Research on underwater image noise reduction method.There are unobvious noise reduction,accentuated image blurring and even elimination of wanted information in the image noise reduction results by using the typical spatial domain and transform domain methods,which being influenced by multiple noises superposition,to process underwater image.Aiming at this problem,this paper has proposed an underwater image noise reduction method via notch filtering and bilateral filtering.The proposed method combine notch filtering,median filtering and bilateral filtering to achieve multiple noises reduction of underwater image.Different from the typical gaussian filtering and median filtering methods which using fixed filter window for noise reduction,the proposed method changes the size of the filter window by estimating the grayscale characteristics of pixels in the neighborhood,and improves the effect of noise reduction by adding a similarity estimation function.Different from the typical wavelet method which zeroing the wavelet coefficients with uniform distribution,the proposed method determines a weighted function by estimating the variance of the adjacent pixels to extract the main frequency components of the noise for facilitating the reduction of image noise.The underwater image noise reduction comparative experiments verifies the effectiveness of the proposed noise reduction method.(2)Research on underwater image segmentation method.for the purpose of meeting the computational time requirement of the intervention for image processing.The authors have used the typical thresholding methods and color-to-gray methods(NTSC and ELSSP methods)with low computational time to process underwater images,because of the low contrast between partial target and background of underwater images,there are inter-regional connection and target partial segmentation of the segmentation results.Aiming at this problem,this paper has proposed an underwater color image segmentation method via RGB channel fusion.The propose method is based on NTSC method and focuses on dynamic estimation of the optimal weights by using the differences of local contrast between RGB channels for RGB channel fusion to obtain the grayscale image with high target-background contrast,and then uses existing thresholding segmentation methods to conduct fast and accurate segmentation.Different from the NTSC method with fixed fusion weights and the ELSSP method without considering the pixels coordinates,the proposed method considers the pixels coordinates and uses the difference among three target-background contrasts of RGB channels to dynamically estimate the fusion weights.Multiple underwater image segmentation comparative experiments verify the effectiveness of the proposed segmentation method.(3)Research on underwater target location method.Considering various distortion,this paper establishes different camera models,and then improves the calibration method proposed by Zhengyou Zhang to calibrate the established camera models.Meantime,this paper conducts model performance comparative experiment to determine the best suitable model for two cameras used in the intervention.The different imaging angle,imaging distance and essential parameters of two cameras causes the difference pixels value of two conresponding images and hard to obtain multiple matching pairs in two conresponding images,so that the typical region matching and feature matching method is not suitable for the research environment.Aiming at this problem,this paper only matches the center point of the target to avoid wrong matching and reduce the computational time.The continuous change of the relative position between two cameras causes the situation where the target is out of the FOV or even repeatedly in and out of the FOV,so that the binocular position measurement method is hard to obtain stable target data.Aiming at this problem,this paper has proposed a target location method combining binocular and monocular vision.The proposed method uses binocular position data to estimate target-size,and adopts monocular position measurement method to obtain the target position data.Underwater comparative experiments verify the effectiveness of the proposed target location method.(4)Development on an underwater target location system based on non-parallel binocular vision.On the basis of the proposed noise reduction method,segmentation method and target location method,this paper has developed an underwater target location system based on non-parallel binocular vision for cooperating the AUVMS to accomplish real underwater autonomous intervention.During the intervention ecperimental research with the cooperation of the developed vision system,there are inter-regional connection and target partial segmentation by using the proposed segmentation method,which being influenced by the poor separation characteristic of target and background in H and I channels during conducting rough segmentation,to process the real underwater intervention images.Aiming at this problem,this paper have improved the proposed segmentation method,and then the underwater segmentation experiments verify the effectiveness of the improved method.The authors have employed the proposed location method to the intervention,because of the particularity of mechanical structure and binocular structure of AUVMS,there are problems that it is difficult to place high-precision calibration boards for the typical calibration method via chessboard to obtain online high-precision zoom calibration results and extrinsic parameter.Aiming at this problem,this paper has proposed a calibration method of location system combining offline and online ways.This method includes two part:offline zoom calibration method and extrinsic calibration method combining offline and online ways.The offline zoom calibration method obtains the required zoom calibration data on the basis of determining the times and degree of camera zooming by multiple trials,and determines the moment of camera zooming during the intervention.The extrinsic calibration method combining offline and online ways utilizes the angle information from the angular transducer mounted on the manipulator joint to obtain the online extrinsic parameter,the underwater calibration experiments verify the effectiveness of the improved calibration method.At last,the underwater autonomous intervention experiment of AUVMS verifies systematically the feasibility of the developed vision system and the proposed noise reduction method,segmentation method and target location method.