Real-time Shrimp Weight Estimation System based on Computer Vision

This repository contains a comprehensive framework for non-invasive, real-time shrimp monitoring. It integrates underwater image calibration, biological data analysis, and advanced deep learning models (YOLOv5-OBB, YOLOv8-Seg, Norfair Tracking) to detect, track, and estimate the weight of shrimp in aquaculture environments.

🌟 Key Features

• Bottom-Up Acquisition Strategy: We utilize a unique camera angle positioned beneath the tank to capture high-contrast ventral profiles. This setup prioritizes shrimp on the tank floor (fixed depth), effectively filtering out floating subjects and surface disturbances. This geometric consistency is critical for minimizing scale deviation and ensuring high-precision biometric estimation.
• Real-time Weight Estimation: Instantly calculates individual shrimp weight (g) from video feeds.
• Water Clarify Model: Automatically detects water turbidity to prevent processing unusable footage, significantly saving power and storage.
• Robust Tracking: Utilizes Norfair + SIFT ReID to track shrimp identities even under occlusion.
• High Precision: Combines Oriented Bounding Boxes (OBB) for detection(Capture the shrimp that is closest to the camera to make sure the precision of length estimation) and Segmentation for precise width measurement.
• Underwater Correction: Includes a dedicated module for correcting optical distortion caused by water.
• Automated Logging: Exports biological data (Length, Width, Weight) to CSV for statistical analysis.

---

📸 Demo

Detection & Tracking & Length, Width, and Weight Estimation

---

⚡ Energy & Storage Optimization

This system is designed for real-world deployment where resources are limited. We implemented two key mechanisms to optimize efficiency:

1. Water Quality Gating (Power Saving)

Before running the computationally expensive detection models (YOLOv5/YOLOv8), the system performs a lightweight check using logistic_water.py.

• Logic: If the water is classified as "Turbid", the system immediately aborts processing for that video.
• Benefit: Prevents wasting GPU/CPU cycles on low-visibility footage, extending battery life for edge devices.

2. Intelligent Data Storage (Storage Saving)

• Turbid Video Handling: Unusable turbid videos are automatically moved to a separate folder and excluded from the analysis pipeline.
• Shrimp-Only Export: The system generates a condensed video containing only the frames where shrimp are detected, filtering out empty frames to drastically reduce file size.

---

📂 Project Workflow & Structure

The project is organized into three main modules. For the best results, please follow the implementation order below:

STEP 1: image_calibration/
      │
      ▼ (Output: Undistorted Images)
      │
STEP 2: Shrimp_size_weight_analysis/
      │
      ▼ (Output: Trained Regression Models .pkl)
      │
STEP 3: shrimp_OBB/
      │
      ▼ (Action: Real-time Inference)
      │
FINAL OUTPUT: Annotated Video & CSV Logs

1. Image Calibration (image_calibration/) Start here. Underwater environments introduce optical distortion. This module calculates the camera matrix and distortion coefficients to correct images before analysis.

• Goal: Minimize geometric error to ensure pixel measurements represent true physical size.

• Key File: undistort_all_img.py

2. Research & Modeling (Shrimp_size_weight_analysis/) The Scientific Foundation. This module analyzes the biological relationship between shrimp dimensions (Length/Width) and Weight.

• Components:

  • linear_relationship: Validates the linearity between pixel and physical measurements.

  • Shrimp_Size_Estimation: Uses LOOCV to train models for converting Pixels → mm.

  • shrimp_weight_regression: Trains the final regression models (Linear/Polynomial) to convert mm → grams.

• Output: Generates the .pkl models used in the real-time application.

3. Real-time Application (shrimp_OBB/) The Deployment. This is the main execution engine that runs on video streams.

• Methodology:

  • Detection: YOLOv5-OBB (Oriented Bounding Box).

  • Refinement: Crop & Rotate -> YOLOv8-Seg (Precise Width).

  • Tracking: Norfair + SIFT ReID.

  • Estimation: Applies the regression models from Step 2 (using image-based pred_length and pred_width).

  • Water clarify: Features logistic_water.py to filter out turbid water, saving computational resources and storage.

---

📊 Experimental Results

Our system has been validated in a real-world aquaculture environment with Penaeus monodon. Key performance metrics include:

• Weight Estimation: The system delivers high-precision biometrics with a weight prediction MAPE of 11.77%. Notably, the discrepancy between our automated estimation and manual ground truth is ≤ 2.2%, demonstrating its capability to replace labor-intensive manual sampling.

• Operational Efficiency:

  - Storage Optimization: By filtering out turbid and empty frames, the system reduced video storage requirements by 69.02% in a month-long deployment.
  
  - Scalability: The pipeline is highly optimized, maintaining a low average GPU utilization of 14.84% (tested on RTX A6000), leaving significant headroom for multi-camera expansion.
  

---

🔗 Acknowledgments & References

Our codebase is built upon the YOLOv5-OBB repository and the Norfair tracking library.

The core integration is implemented within shrimp_OBB/detect_norfair_optimize.py, where we added Norfair tracking with SIFT Re-Identification (ReID) and embedded the regression logic to predict the actual length, width, and weight of the shrimp in shrimp_OBB/utils/plots.py.

Researchers interested in the underlying methodology can utilize or modify to the code provided in this repository.

Full research paper will be updated once it is accepted.

---

Author

Bing Chian 
email: cucuchen105@gmail.com