🐘 Automated Wildlife Tracking & Re-Identification Pipeline

November 25, 2025 · View on GitHub

Paper

This repository contains the official code and implementation for the paper "A framework for detecting and tracking elephants in drone videos", published in Drone Systems and Applications (2025).

It provides an end-to-end computer vision pipeline designed to detect, track, and re-identify animals in drone video footage. While developed specifically for tracking Elephants, the framework is adaptable to other species.

It solves the common problem of "ID Switching" (where a tracker loses an animal and re-assigns a new ID) by using a novel post-processing algorithm based on visual similarity (SIFT/SSIM) and physical feasibility.

🔗 Read the Full Paper

🌟 Features

  • Detection & Tracking: Uses YOLOv11 combined with BotSort for robust frame-by-frame tracking.
  • Smart Re-Identification: automatically stitches broken tracks together using:
    • SIFT Features & Homography (to align animal bodies across frames).
    • SSIM (Structural Similarity) (to compare visual identity).
    • Velocity Checks (to ensure merges are physically possible).
  • Analytics Dashboard: Automatically generates:
    • Movement trajectories (X/Y plots).
    • Distance traveled per ID.
    • Herd density heatmaps.
    • Social interaction (overlap) statistics.
    • Cluster analysis (identifying sub-groups within the herd).
  • Visualization: Outputs a fully annotated video with corrected IDs.

📂 Repository Structure

├── notebooks/
│   ├── TrainingModel.ipynb     # Train YOLOv11 on your custom dataset
│   ├── EvaluateModel.ipynb   # Calc Precision, Recall, mAP & Confusion Matrix
│   └── ElephantTrackingPipeline.ipynb    # Main Pipeline: Tracking -> ReID -> Viz -> Stats
├── configs/
│   ├── botsortV4.yaml                # Tracker configuration
│   ├── TestData.yaml                 # Dataset configuration
├── scripts/
│   ├── Viz.py                        # Analytics generation script
│   └── process_pipeline.py           # Python script version of the pipeline
└── README.md

🧠 Model Weights

To run this pipeline, you need the trained YOLOv11 model weights used in our study.

📥 Download Trained Weights (best_xl.pt)

Place the downloaded best_xl.pt file in the root directory or your Google Drive working folder.

🚀 Getting Started

Prerequisites

The project is optimized to run on Google Colab (for GPU access) or a local machine with a dedicated GPU.

Install Dependencies:

pip install ultralytics opencv-python pandas numpy scikit-image seaborn tqdm matplotlib

Running the Pipeline

  1. Training (Optional): If you want to train the model from scratch on your own dataset, open notebooks/1_Training_Pipeline.ipynb. Ensure your dataset is formatted according to YOLO standards.

  2. Evaluation: To test the model's accuracy (mAP, Precision, Recall), run notebooks/2_Evaluation_Pipeline.ipynb.

  3. Inference (Tracking & Analysis): This is the main tool. Open notebooks/3_Inference_Pipeline.ipynb.

    • Set INPUT_VIDEO to your video path.
    • Set MODEL_PATH to the downloaded best_xl.pt.
    • Run all cells to generate videos and statistics.

📊 Outputs

The pipeline automatically creates a pipeline_results folder containing:

  • Final_Output_ReID.mp4: The video with corrected tracking IDs overlaid.
  • tracking_processed.csv: Cleaned tracking data after Re-ID.
  • stats/:
    • id_statistics.txt: Seconds/Frames visible per ID.
    • overlap.csv: Interaction percentages between animals.
    • cluster_frame_ranges.csv: Grouping dynamics over time.
    • sum_distance_per_id.csv: Total distance traveled.
  • plots/:
    • trajectories.png: Path of movement for every animal.
    • density_heatmap.png: Where the herd spent the most time.
    • avg_distance_plot.png: Average movement speed of the herd.

📚 Citation

If you use this code or model in your research, please cite our paper:

Chaim Chai Elchik, Serge Wich, and André Burger. 2025. A framework for detecting and tracking elephants in drone videos. Drone Systems and Applications. 13: 1-16. https://doi.org/10.1139/dsa-2025-0032

📜 License

This project is open-source and available under the MIT License.