Fish fry body length measurement with improved YOLOv8n-pose and biomechanics

Accurate measurement of fish fry body length is crucial in biomechanical research and the development of intelligent aquaculture, as it directly affects the growth, locomotion, and ecological adaptability of fish. Traditional manual methods are time-consuming, labor-intensive, and may harm fish fry. Therefore, accurate, rapid, and non-destructive measurements of large quantities of fish fry are highly important in aquaculture. This study used 20–100 mm grass carp fry (Ctenopharyngodon idella) as test subjects. An image acquisition platform was developed to obtain RGB-D data from the top view of the fry. We proposed ROS-YOLO, which replaces the original C2f module of YOLOv8n-Pose with reparameterized convolution-based shuffle one-shot aggregation (RCS-OSA) and introduces a simple attention module (SimAM) into the main feature extraction layer, to detect key body length points of fish fry. Depth information for 3D keypoint coordinate transformation was obtained through the depth map. Additionally, biomechanical principles were incorporated to study the movement patterns, muscle activity, and hydrodynamic efficiency of fish fry. High-speed cameras and motion tracking software were used to analyze swimming kinematics and dynamics, while biomechanical modeling was employed to simulate the effects of water flow on growth and development. Finally, fish fry body lengths were calculated based on keypoint coordinates. In experiments, ROS-YOLO achieved an average keypoint detection accuracy of 99.2%, with 3.97 M parameters and 125 FPS. Compared to manual measurements, the overall average error in automatic measurement results was 2.87 mm (5.85%). Therefore, the proposed method meets real-time measurement requirements for fish fry body length and provides insights into the biomechanics of fish fry growth and movement.