TL;DR
This paper introduces a novel computer vision framework (YOLOv8 + ByteTrack + Kalman filtering) to precisely track shrimp behavior in recirculating aquaculture systems. They discovered that a moderate water flow rate (5.0 cycles/day) creates a unique "hydrodynamic window" where shrimp exhibit significantly different movement patterns during feeding versus non-feeding periods, crucial for optimizing feeding strategies.
What Problem Does This Paper Address?
Optimizing aquaculture efficiency, particularly for shrimp like Litopenaeus vannamei, hinges on precise control of environmental factors. Water velocity is a key one, influencing everything from feeding behavior to energy expenditure. However, quantitatively understanding how water velocity impacts shrimp behavior in complex, dynamic recirculating aquaculture systems (RAS) has been challenging due to limitations in tracking individual animals and analyzing their movements over time. Existing methods struggle with occlusions, high densities, and the need for precise kinematic data. This paper aims to bridge this gap by developing a robust system to accurately track shrimp and analyze their behavior under varying water flow conditions.
Methods
The authors developed a multi-objective behavioral analysis framework that integrates three core components:
- Object Detection: They employed the YOLOv8 model, known for its effectiveness in detecting small and occluded objects, which is crucial for shrimp in RAS environments.
- Multi-Object Tracking: ByteTrack was used for its robust tracking capabilities, especially in crowded scenarios. It utilizes a dual-threshold matching strategy to maintain trajectory continuity even when objects are temporarily occluded or have low detection scores.
- Behavioral Interpretation & Correction: Kalman filtering was integrated to correct for coordinate offsets caused by water refraction, a common issue in underwater imaging. This ensures accurate extraction of kinematic parameters like velocity and turning angle.
The system was tested under three controlled water circulation rates (2.0, 5.0, and 10.0 cycles/day) using high-frequency (30 fps) video recordings. They specifically analyzed movement velocity and turning angle during both feeding and non-feeding periods.
Datasets
The study utilized experimental data collected from a custom-built recirculating aquaculture system. They collected videos of Litopenaeus vannamei under different water circulation rates. For training and evaluating the object detection model (YOLOv8), they created an annotated dataset using the PASCAL VOC standard format, with bounding boxes and categories for shrimp.
Key Insights and Ideas
- The "Hydrodynamic Window": The most striking finding is the identification of a specific water flow rate (5.0 cycles/day) that creates a distinct behavioral difference between feeding and non-feeding states. This suggests a critical range where shrimp prioritize feeding efficiency without expending excessive energy on movement.
- Velocity vs. Turning Angle: While water velocity significantly impacts shrimp movement speed, it has a negligible effect on their turning angle. This implies that flow primarily influences the pace of movement rather than the directionality of turns.
- Crustacean vs. Fish Behavior: The study highlights that Litopenaeus vannamei exhibits a different flow adaptation strategy compared to fish. They appear to suppress movement during feeding at moderate flows to conserve energy, contrasting with fish that might increase metabolism.
- Integrated Vision System: The combination of YOLOv8 for detection, ByteTrack for tracking, and Kalman filtering for correction provides a powerful, end-to-end solution for analyzing animal behavior in challenging aquatic environments.
Key Results
- The YOLOv8 model achieved high performance metrics (Precision: 90.0%, Recall: 87.6%, F1-score: 88.3%, mAP50: 92.8%), outperforming YOLOv5 and Faster R-CNN.
- Water circulation rates significantly affected shrimp movement velocity.
- A statistically significant difference (p < 0.05) in average movement velocity was observed only at the moderate circulation rate of 5.0 cycles/day, where feeding velocity was significantly lower than non-feeding velocity (13.58 vs. 23.57 pixels/s, a 42.4% reduction).
- Turning angle remained consistent across all tested circulation rates and feeding states, showing no statistically significant differences.
Limitations and Future Work
The authors acknowledge several limitations:
- Turbidity and Surface Waves: The system's performance can degrade in highly turbid water or with significant surface waves, which can introduce spatial errors.
- Long-Term Occlusions: While ByteTrack handles short-term occlusions well, irreversible ID switches can occur during prolonged periods (>15 seconds) of complete occlusion, especially in high-density situations.
- Physiological Validation: The study lacks direct physiological measurements (e.g., stress hormones) to definitively link observed behavioral shifts to adaptation or stress responses.
- Environmental Factor Interactions: The influence of other environmental factors (like dissolved oxygen) on the observed behaviors was not deeply explored.
Future work proposed includes:
- Enhancing robustness with near-infrared imaging, integrating ByteTrack's re-identification capabilities, and developing turbulence-adaptive CNNs for refraction correction.
- Linking behavioral data with biomarkers and exploring couplings with other environmental factors.
- Integrating the framework with Spatiotemporal Graph Convolutional Networks (ST-GCNs) and real-time biosensor data to create closed-loop systems for precise feeding and flow regulation.
Summary updated