Algorithm for automatic counting of fish in an image and tracking their movement based on the YOLOv9t neural model

Purpose of research. Traditional methods based on visual observation and manual counting not only have obvious limitations in terms of time and human resource costs but also yield insufficiently accurate results due to the subjective human factor involved in the process. These inaccuracies, even minor ones, can lead to erroneous management decisions, which negatively impact production efficiency in aquaculture. Methods. To eliminate these shortcomings, this paper presents an automated solution that utilizes the YOLOv9t neural network model for the task of detecting and counting fish in underwater images. Thanks to the optimized architecture of the YOLOv9t neural model, which contains only 2 million parameters, it demonstrated high performance in identifying fish in images from the DeepFish dataset, with the following evaluation metrics: Precision - 0.928, Recall - 0.91, mAP50 - 0.961, and mAP50-95 - 0.584. The Non-Maximum Suppression method was used to eliminate duplicate detections of fish in the same area, while the application of the DeepSORT algorithm enabled the continuous tracking of each individual across video frame sequences by assigning unique identifiers. Results. The research results confirmed that the YOLOv9t neural model is suitable for creating automated video analytics systems in aquaculture for monitoring fish behavior and managing activation devices. This enables the transition of key control processes to a fully automated basis, thereby optimizing resource utilization. The proposed architecture provided high accuracy and reliability across various environmental conditions-from clear to murky wateropening prospects for application in real-world production environments. Conclusion. This operational stability makes the system ready for industrial-scale implementation with the aim of enhancing farm management efficiency.