Abstract: Aiming at the complex experimental environment of waterway hydraulic models characterized by high impurity content and strong interference,this study constructed an image threshold segmentation method (including low-threshold and high-threshold conditions) and a deep learning method based on the YOLOv5 model for fish behavior detection.The detection performance of the two object detection methods for individual and group fish behavior was compared from three perspectives:visual effects,discrepancy ratio of detection results,and fish swimming trajectories.Results demonstrate that the deep learning method significantly outperforms the image threshold segmentation method in the hydraulic model experimental environment,as evidenced by both qualitative description and quantitative analysis.The deep learning approach exhibits higher target recognition accuracy,stronger robustness against environmental impurities,and smoother reconstructed fish swimming trajectories.This study further analyzed the advantages and disadvantages of the two methods and their applicability to different experimental scenarios from the perspectives of operational workflow,application maturity,target universality,limiting factors,multi-class detection capability,and application prospects.These findings can provide references for the selection of fish behavior detection methods and the optimization of detection performance in waterway hydraulic model experiments.