Detection of fish feeding behavior based on optical flow and schooling behavior theory

To analyze multi-dimensional, dynamic fish behavior during feeding, this paper proposes a novel framework that integrates multi-source information for quantifying feeding desire. The core of this framework is to leverage synergistically both the average swimming speed of fish schools, which is extracted by using optical flow techniques, and the multi-level spatial structure indicators (comprehensive aggregation index and convex hull compactness), which constructed by schooling behavior theory, followed by a quantitative classification approach to achieve the classification of feeding desire. To support this framework, this paper embeds PPA and SEAM modules into YOLOv8, and the detection accuracy reached 96.68% on a self-constructed dataset, which provided a solid foundation for upstream feature extraction. Experimental results show that this framework achieved an accuracy of 90.85% on the classification of feeding desire.