Abstract
Ammonia nitrogen, as a “deadly killer” in aquaculture, poses a fatal threat to fish due to its accumulated concentration. Fish usually exhibit different behavior characteristics before and after being subjected to ammonia nitrogen stress. However, the traditional behavior analysis methods have the disadvantages of easy detachment and interfere with the experimental results. To address this, we propose multiple object detection and tracking models under different conditions to enable both qualitative and quantitative analysis of ammonia nitrogen stress responses in fish. By integrating high-resolution cameras and deep learning algorithms, these methods automatically analyzes behavioral indicators (e.g., swimming distance, velocity, spatial distribution, etc) correlated with ammonia exposure. Experiments were conducted on bass, sturgeon, and carp under controlled ammonia nitrogen concentrations, with video data processed using convolutional neural networks (CNNs) for feature extraction. Furthermore, by optimizing camera configurations (quantity and placement), we developed a 3D reconstruction method for fish locomotion to enable more precise behavioral analysis. Results demonstrated over 90% accuracy in fish detection, significantly outperforming traditional manual observation. The methods provide a cost-effective, scalable solution for early stress detection, aiding precision aquaculture management. This work highlights the potential of computer vision in transforming water quality monitoring and improving fish welfare.