Research On Automatic Measurement Technology Of Underwater Seafood Based On Binocular Vision

At present,the fishing of marine products mainly depends on artificial,which is inefficient,and the underwater environment will cause great harm to divers' health.Therefore,it has become a trend to use underwater fishing robot to replace artificial fishing.Usually,when the underwater fishing robot is fishing for seafood,it needs to find the seafood,and then select the mature seafood to grab,which requires the robot not only to detect the seafood,but also to judge its size.In this paper,sea cucumber is taken as the main research object,and the underwater image is obtained by binocular camera.According to the characteristics of underwater environment,the precise measurement of binocular and the detection of seafood are researched.On this basis,an automatic measurement method of seafood size based on binocular vision is designed,which is used to find and judge the size of the seafood.The main contents of this paper include the following parts:(1)Underwater binocular vision technology was researched.The 3D information obtained by binocular is not accurate due to the refraction phenomenon.In this paper,through the underwater camera imaging model,it is found that refraction causes a great influence to camera focal length and imaging distortion.Then the underwater binocular camera is calibrated according to the derived results.On this basis,the binocular camera model is built and the actual position of the target point on the image in space is deduced.Finally,through the underwater 3D reconstruction and the underwater 3D measurement experiment,We found that the average error of underwater reconstruction results is less than 0.154mm,and the average error of adjacent corner measurement is less than 1.29%.(2)The detection methods of underwater products was researched.In view of the demand for accuracy and speed of the target detection;algorithm,this paper uses the yolov3 target detection algorithm to identify holothurian and locate the region of interest,and analyzes the detection principle and regional positioning principle.In order to train the holothurian detection model,this paper uses the underwater robot to collect a large number of underwater holothurian pictures for annotation,and combines the competition dataset published by URPC-2019 to form the underwater holothurian data set.In this paper,K-means algorithm is used to cluster the annotation boxes of dataset to get the new anchor boxes.In this paper,the holothurian dataset and news anchor boxes is used to train the detection model,the model detection mAP of the model is 90.46%.Through the relationship between the threshold and F1 score,we finally selects 0.5 as the threshold,and the final recall rate reaches 80.34%.(3)The automatic measurement method of underwater product size was researched.Firstly,the underwater binocular camera is used to obtain the corrected image and scene depth information;the YOLOv3 detection model is used to find the bounding boxes of holothurian;Then,a novel GrabCut-RGBD image segmentation method is constructed based on the Gaussian model of fusion color and depth information,and regions of interest with sea cucumber are separated from the back-ground in the bounding box;by using the minimum adjacent rectangle algorithm,th e optimal size measurement points are found.Through the Euclidean distance,the measurement result of holothurian is calculated,and the final automatic measurement error is less than 4.85%.