Research On The Method Of Fish Target State Estimation Based On Computer Vision

With the rapid development of deep neural network and underwater binocular vision technology,fish farming is transforming to scientific and intelligent direction.Computer vision methods have important application value in monitoring the survival status of fish targets due to their high accuracy,fast speed,and non-contact characteristics.In this paper,the underwater fish target is the research object,the size and movement speed of the underwater fish target are used to evaluate the survival state of the fish,and the fish target state estimation method based on computer vision is studied.The main research contents of this paper are as follows:In response to the actual needs of fish breeding in marine cabins,the underwater fish target detection method is studied,and the Yolov3 algorithm that meets real-time requirements and has a high accuracy rate is used to extract underwater fish targets,and build a fish target detection data set,Use the K-means clustering algorithm to form a new a priori box,and realize the fish target trajectory tracking through the associated Deepsort multi-target tracking algorithm.Aiming at the refraction phenomenon and the imaging distortion caused by the binocular camera when used underwater,the parallel binocular camera model and the underwater camera imaging model are established,and the underwater camera calibration experiment is carried out to obtain the camera’s internal reference data and the relative position of the binocular Relations,binocular underwater distortion correction,combined with the SGBM stereo matching algorithm for three-dimensional reconstruction of underwater fish targets.Research on fish target state estimation method based on computer vision.Through the Yolov3 fish school target detection algorithm,the two-dimensional and three-dimensional conversion relationship of underwater coordinate points and the Euclidean distance formula,the size of the fish target is estimated.On this basis,the Deepsort multi-target tracking algorithm is combined to realize the trajectory tracking of the fish target,and then realize the speed estimation of the fish target,and finally realize the monitoring of the fish’s survival state.Underwater fish target size estimation and velocity estimation experiments show that the average relative error of fish size estimation when the camera is placed in the water is 3.57%,and the average relative error of fish lateral movement velocity estimation under natural light conditions is 1.60 %.