Research progress of fish video 

tracking application based on computer vision

Abstract: Object tracking technology is an important research direction in the field of computer vision. The introduction of computer vision technology in the fields of marine environment detection and aquaculture monitoring can realize the tracking of fish video targets, which can save unnecessary manpower and material resources, in aquaculture, marine environment monitoring and other fields, it can reduce costs and improve efficiency. At present, there has been a lot of research in the field of underwater image vision, but there is a lack of general summary of the research status of video tracking technology for fish targets. Therefore, a comprehensive review of fish video tracking methods based on computer vision techniques is of great significance. Fish video tracking is mainly divided into four parts: underwater image acquisition, image sharpening, fish tracking and trajectory output. Among them, the image sharpening part and the fish target tracking part are the most critical links in the overall process. This article makes a comprehensive review of these two parts.Underwater images captured in natural conditions are affected by the refraction of light in water, and different wavelengths of light exhibit different degrees of exponential decay underwater, resulting in blurring, color cast, and reduced visibility in the captured underwater images. Underwater image sharpening technology can be used to solve these problems. The research directions in this field are mainly divided into image enhancement, image restoration and deep learning. The image enhancement technology realizes the sharpening of the image by adjusting the color of the underwater photographed image. There are mainly methods based on the histogram stretching method and the method based on the Retinex theory. The method based on histogram stretching has high operating efficiency, but the application range is narrow and noise is easily introduced; the method based on Retinex theory has better effect on image color correction and edge sharpening, but the operation is more complicated and the execution efficiency of the algorithm is relatively low. Image restoration technology achieves image clarity processing by establishing a degradation model of underwater images. This method has a significant effect in certain situations, but has poor applicability in complex scenes. The deep learning method realizes the clear processing of underwater images by learning the features between blurred underwater images and clear images. This method has strong applicability, and the effect of color restoration is remarkable, but the phenomenon of blurred details and unclear edges will occur.Underwater fish video tracking is mainly affected by the uncertainty of the motion state of fish target and the uncertainty of observation data. In the field of fish target tracking, according to different types of observation models, it is mainly divided into generative methods and discriminative methods. The generative method realizes the tracking task of the target by analyzing the target features in the first frame of the video image, generating a tracking template, and searching for the target closest to the template in the subsequent image frames. The generative method is relatively simple to implement and has high computational efficiency, but the tracking accuracy decreases when the shape of the fish target changes or is occluded by obstacles. The discriminative method transforms the target tracking problem into a classification problem, and uses the classifier to distinguish the fish target and the background, so as to further realize the tracking of fish target. The main research directions of discriminative methods are divided into correlationbased filtering methods and deep learning methods. The basic idea of the target tracking method based on correlation filtering is to use a preset filtering template to perform convolution operation on the template in the next frame of image and calculate the response value. The area with the largest response value is where the fish target is located. In recent years, deep learning methods have outstanding performance in the field of object classification and are suitable as classifiers in discriminative tracking methods. In contrast, the correlationbased filtering method has the effect of antideformation and antiocclusion, and has high operating efficiency, but requires a preset filtering template and has poor applicability; the deep learning method has relatively high accuracy for target detection and tracking, but a large amount of image and video data is required for model training, and the efficiency of model training is low.To summarize the full text, the development of computer vision technology provides a new observational approach for underwater fish behavior analysis and ecosystem monitoring. Through underwater monitoring, images of fish behavior can be obtained in real time, which can intuitively reflect the survival status of fish and the richness of ecosystems. Using computer vision technology to process underwater surveillance video can efficiently and cheaply obtain fish and ecological environment information, and provide a reference for the management and assessment of marine fishery resources. However, computer vision technology still has certain deficiencies and limitations in underwater fish video tracking scenarios. For example, due to the influence of illumination and hydrological conditions, the phenomenon of light scattering under water is serious, and the observation coverage of video surveillance equipment is limited. Better image enhancement or image restoration methods are needed to clarify underwater images to improve the detection efficiency and tracking accuracy of underwater fish targets. Among the many methods for realizing fish video tracking, traditional methods have complete theory and mature algorithms, but their adaptability is limited in special environments; on the one hand, deep learning methods have wider applicability and higher accuracy, but the model training time is longer, and more resources are occupied. On the other hand, the model is relatively complex, and it is necessary to compress the model to reduce the occupancy of hardware resources by the model, so that it can be easily transplanted into embedded devices. The method of tracking underwater fish targets with computer vision technology has more and more prominent advantages under the general trend of fishery resource survey and automatic processing of marine ecosystem monitoring, and is becoming the main development direction in the future.

Keywords: computer vision; fish video; underwater image sharpening; fish tracking