Phase 1: Initiating the Cow Detection Project

2024-11-04

  • The project began with the goal of identifying individual cows based on their unique features, such as blue spray-painted dots.
  • Image data was organized into directories containing side and back views of cows.
  • A dataset of annotated images was created using SAM (Segment Anything Model) for segmentation and labelling.

2024-11-12

  • Annotated images were further refined using SAM to improve segmentation quality.
  • Initial exploration of YOLO (You Only Look Once) for object detection began, focusing on individual cow identification.

2024-11-20

  • Validation and training datasets were finalised.
  • Initial YOLOv5 training commenced. Training steps, challenges, and solutions were logged for documentation purposes.
  • Early metrics (recall and mAP) highlighted the need for additional annotations and hyperparameter tuning.

Phase 2: Scaling the Dataset and Enhancing Model Accuracy

2024-12-05

  • Dataset expanded to include 72 images categorized into “back view” and “side view.” Measurements from an Excel file were linked to individual images for further analysis.
  • Exploratory work began on feature-based identification using sprayed dots as unique identifiers.

2024-12-09

  • Focus shifted to YOLOv8 for its improved speed and accuracy.
  • A large dataset of 29,000 images was prepared. Data augmentation, hyperparameter tuning, and extended training epochs were implemented to improve accuracy.
  • Training and validation processes were fine-tuned, with ongoing monitoring of model performance.
  • Plans for video recognition using Deep SORT for tracking were initiated.

Phase 3: Introducing Behaviour Detection

2025-01-05

  • The project expanded to include cattle behaviour detection, with bounding boxes labelled for grazing, lying, and standing behaviours.
  • Training datasets were organised
  • Validation datasets were structured to ensure accurate testing of behaviour detection models.
  • Training commenced with a YOLOv8 model, using default hyperparameters and an L4 GPU for faster processing.

2025-01-07

  • After initial training, results were visualized to ensure proper bounding box placement and behaviour categorization.
  • Overfitting was monitored closely, and early results suggested adjustments in label consistency and data augmentation strategies.

Phase 4: Transitioning to Video Detection and Deep SORT Integration

2025-01-08

  • Integration of Deep SORT for tracking individual cows in video feeds began.
  • Preliminary testing involved tracking behaviours such as grazing and lying across consecutive frames.
  • Video data from sheds was used to validate the model’s robustness in real-world scenarios.

Phase 5: App Deployment and Real-World Applications

2025-02-01 (Planned)

  • Begin development of a mobile app to integrate cow detection and behaviour monitoring models.
  • The app will leverage mobile phone cameras for real-time detection and integrate QR code-based identification for individual cows.
  • Sustainability features, such as logging health events, reproductive cycles, and behaviours, will be included.
  • Integration with farm management systems will ensure seamless adoption in real-world settings.

2025-03-15 (Planned)

  • Deployment of the app in pilot farms for testing and feedback.
  • Real-time monitoring of cattle behaviours and health indicators will be assessed for scalability.

Summary and Future Plans

The cow detection project has evolved significantly since its inception, progressing from basic image identification to video tracking and behaviour analysis. The integration of AI models such as YOLOv8 and Deep SORT has enabled precise detection and tracking, while the planned app deployment will make these tools accessible to farmers, contributing to sustainable and efficient livestock management.

Key milestones include: - Transitioning from static image detection to video-based tracking. - Expanding the scope to include behaviour detection for welfare and productivity improvements. - Designing a user-friendly mobile app to bring AI-powered cattle management to farms.

Future Goals: - Scale the solution to monitor larger herds and diverse farming environments. - Incorporate advanced AI features, such as stress detection and predictive health analytics. - Collaborate with stakeholders to enhance adoption and refine app functionality based on user feedback. ““”