Research On The Method Of Fish School Video Tracking

China is a major maritime country with abundant marine resources.In the process of marine resource development,marine engineering activities such as shipping,undersea tunnels,offshore bridges and offshore wind farm construction will generate a lot of man-made underwater noise.The changes in the underwater acoustic environment caused by underwater noise will have a significant impact on the life activities and health of fish schools.When studying the impact of underwater noise on fish schools,most of them are observed and recorded by human eyes and recorded obvious abnormal behaviors such as jumping out of the water under sound stimulation,but failed to quantitatively analyze the impact of noise interference on fish schools.There is not enough attention to abnormal behaviors caused by underwater noises,such as motion trajectories,speed changes,etc.,which are not obvious.Monitoring the changes in the trajectory of a school of fish under acoustic stimulation,including movement parameters such as sudden changes in direction and changes in speed,is an effective way to understand the degree of influence of the school of fish on different acoustic stimulation intensities.The acquisition of these motion parameters is based on accurately tracking the trajectory of the fish school.The use of computer vision to process fish school videos to obtain fish trajectories has been applied,and the processing process is accurate and efficient,but the existing fish school monitoring systems are mostly proposed for specific fish,and it is not easy to migrate to other types of fish tracking.This paper focuses on the small and medium-sized fish schools that are common in scientific research and aquaculture.First,research on fish school video trajectory tracking from the perspective of being simple and easy to implement and transferable to different fish species;then deep learning is applied to fish school video trajectory tracking to achieve high accuracy,robustness and real-time.The main research contents and results are as follows:1.In view of the problem that the existing fish school video trajectory tracking systems are mostly proposed for the gray-scale features of the body surface of specific fish such as zebrafish,and it is difficult to migrate to other fish tracking,this paper proposes a fish school video based on motion features.Trajectory tracking algorithm.When the fish swims in a two-dimensional plane,the area of the fish’s top-view angle remains almost unchanged,and the fish’s swimming direction is unlikely to change in a short time(such as two consecutive frames).These two characteristics are combined with the nearest neighbor method.Group tracking.And use the method of image corrosion to solve the problem of mutual occlusion between fish bodies.When the proposed algorithm was applied to track the trajectory of fish school videos in two scenes of different complexity including 10 small crucian carp fry and 11 large yellow croaker about 7 months old,the tracking accuracy reached 95%and 90%,respectively.Above,it is verified that the proposed algorithm can be applied to the trajectory tracking of different fish swimming.The thesis adopts a deep learning-based target detection-data association fish school video trajectory tracking method,and for the first time the deep learning method is applied to the data association part of the fish school video trajectory tracking research.2.The tracking accuracy of the tracking method based on motion characteristics is not high enough when applied to more complex scenes.In order to solve the above problems,a deep learning method is used to quickly and accurately track the trajectory of the fish school.Deep learning has excellent performance in the field of target detection and target tracking.It has the characteristics of high accuracy,real-time and avoiding manual feature extraction.First of all,the large yellow croaker and crucian fish swimming image data sets were established respectively,and the detection effects of the two target detection models,YOLOv5 and SSD,after training on the data set were compared.In the end,the YOLOv5 algorithm with higher detection accuracy and faster detection speed was selected to complete the target detection part and obtain the fish position information.In order to improve the detection accuracy,the data set is expanded by the method of data enhancement,which provides reliable input for the data association part.In the data association part,the DeepSORT algorithm is used to associate the fish position information detected by YOLOv5 on the time axis to obtain the fish trajectory.When the proposed algorithm is applied to the video tracking of fish schools in two scenes with different complexity,including 10 small crucian carp fry and 11 large yellow croaker about 7 months old,the tracking accuracy reaches 98%and 94%respectively.For complex problems such as water surface reflection,underwater rocks and fish body occlusion,it has a good tracking effect,and the average single frame processing time is only 0.014s.Experimental results show that the method based on deep learning can be used for trajectory tracking of fish school videos,and compared with the proposed method for tracking fish school video trajectories based on motion features,it has improved tracking accuracy and robustness.Single frame processing time can meet real-time requirements.3.In addition,in order to facilitate the use of relevant researchers,especially engineering application personnel,based on the Python language and PyQt development platform,a front-end graphical user interface for fish school video trajectory tracking is made,which can quickly realize fish school video detection and tracking,as well as fish school picture detection,etc.Features.