This repository contains the code, models, and data used in our paper:

Tracking Moose Using Aerial Object Detection
Christopher Indris, Raiyan Rahman, Goetz Bramesfeld, Guanghui Wang
Toronto Metropolitan University, 2025

We explore aerial wildlife tracking with a focus on detecting moose in snowy environments using object detection models onboard drones. The key challenge lies in spotting small objects in high-resolution aerial imagery under strict computational limits.

To address this, we:

• Propose a multi-scale patching method to reduce input resolution while maintaining high detection performance.
• Compare YOLOv11, Faster R-CNN, and Co-DETR across different patching configurations.
• Analyze model performance under varying threshold and overlap parameters.

The result: YOLOv11, despite being the lightest model, achieves comparable or better performance than the others—making it well-suited for UAV deployment.

• Dataset: 1694 high-resolution aerial images of moose, divided into patches of varying sizes (large, medium, small). Available Soon!
• Models: YOLOv11, Faster R-CNN, Co-DETR evaluated across a grid of patching hyperparameters.
• Best Result: All models reached ≥93% mAP@IoU=0.5 under at least one configuration.
• Efficiency: YOLOv11 achieved 93.2% [email protected] using only 2.6M parameters and ~0.17 sec/iter.

• 📁 Notebooks/ --> Python runs
• 📁 R/ --> analysis

We use PyTorch, Ultralytics YOLO, and MMDetection, and R for analysis. See the .ipynb notebooks for installation details.

If you find this work useful, please consider citing:

@misc{indris2025trackingmooseusingaerial,
      title={Tracking Moose using Aerial Object Detection}, 
      author={Christopher Indris and Raiyan Rahman and Goetz Bramesfeld and Guanghui Wang},
      year={2025},
      eprint={2507.21256},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.21256}, 
}

This project was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) under grants ALLRP 588173-23 and ALLRP 570580-21.