PineSORT: A Simple Online Real-time Tracking Framework for Drone Videos in Agriculture

Danny Xie-Li¹   ·   Fabian Fallas-Moya^{2    
1} Computer Science Department, Instituto Tecnologico de Costa Rica, Cartago, Costa Rica  
² Atlantic Campus, Universidad de Costa Rica, Cartago, Costa Rica  

We introduce PineSORT, a novel Multiple Object Tracking (MOT) system for drone-based agricultural monitoring, specifically tracking pineapples for yield estimation. Our approach tackles key challenges such as repetitive patterns, similar object appearances, low frame rates, and drone motion effects. PineSORT enhances tracking accuracy with motion direction cost, camera motion compensation, a three-stage association strategy, and overlap management. To handle large displacements, we propose an ORB-based camera compensation technique that significantly improves Association Accuracy (AssA). Evaluated via 5-fold cross-validation against BoTSORT and AgriSORT, PineSORT achieves statistically significant gains in Identity-Switch Penalized IDF1 (ISP-IDF1), IDF1, HOTA, and AssA metrics, confirming its effectiveness for tracking low-FPS drone footage and its value for precision agriculture.

---

📁 Project Structure

src/
├── main.py                   # Main entry point to run PineSORT
├── requirements.txt          # Python dependencies
├── output/                   # Output folder for tracking results
│   └── track/
│       └── tracker/
│           ├── debug/        # Debugging info and logs
│           └── images/       # Visualizations of tracking outputs
├── output.txt                # Summary output file
└── tracker/
    ├── gmc.py                # Global motion compensation utilities
    ├── matching.py           # Object matching and association logic
    ├── pinesort.py           # Core PineSORT tracker implementation
    ├── track.py              # Tracking management code
    └── utils.py              # Helper functions

💻 Installation

1. Clone this repository:

   git clone https://github.com/yourusername/pinesort.git
   cd pinesort/src

2. Install dependencies:

   pip install -r requirements.txt

⚙️ Usage

Run PineSORT on a video dataset by specifying parameters as needed. Below are the available command-line arguments with their defaults and descriptions:

Dataset and Evaluation Arguments

• --path_video_imgs (str): Path to the folder containing video frames as images.

Detector

• --model_path (str): Path to YOLO object detector weights (Ultralytics YOLO).

PineSORT Tracking Parameters

Parameter	Type	Default	Description
--det_thresh	float	0.40	Detection confidence threshold.
--min_det_thresh	float	0.30	Minimum detection threshold to consider.
--max_age	int	5	Max frames to keep alive a track without detection.
--min_hits	int	1	Min hits before confirming a track.
--first_iou_threshold	float	0.30	First IOU threshold for association stage.
--second_iou_threshold	float	0.30	Second IOU threshold for association.
--third_iou_threshold	float	0.10	Third IOU threshold for association.
--overlap_iou_threshold	float	0.10	IOU threshold to suppress redundant boxes using Kalman predictions.
--delta_t	int	3	Time delta for motion model update.
--inertia	float	0.2	Inertia coefficient for motion modeling.
--asso_func	str	"eiou"	Association function (e.g., "iou", "giou", "eiou", "ciou", "diou").
--use_byte	boolean	False	Enable ByteTrack-style association logic (flag).
--img_width	int	1920	Image width in pixels.
--img_height	int	1088	Image height in pixels.
--camera_compensation	str	"orb"	Camera compensation technique (e.g., "orb", "optisparseOptFlow", "sift").

Tracking Arguments

Parameter	Type	Default	Description
--track_high_thresh	float	0.3	Confidence threshold for tracking.
--track_low_thresh	float	0.1	Lowest detection threshold valid for tracks.
--new_track_thresh	float	0.7	Threshold to start a new track.
--track_buffer	int	3	Number of frames to keep lost tracks before deletion.
--match_thresh	float	0.7	Matching threshold used during tracking association.
--aspect_ratio_thresh	float	1.6	Filter out boxes with aspect ratio above this value.
--min_box_area	float	10	Filter out tiny bounding boxes below this area.

Output Arguments

• --output (str): Directory to save tracking results (MOT format, debug info, images). Default: ./output/track/
• --output-debug (flag): Save debug images with detector and Kalman predictions. Default: enabled (True)
• --output-frames (flag): Save images with tracking IDs overlaid. Default: enabled (True)

---

Example command

python main.py \
  --path_video_imgs "./images" \
  --model_path "./best.pt" \
  --det_thresh 0.4 \
  --max_age 5 \
  --asso_func eiou \
  --camera_compensation orb \
  --output "./output/track/" \
  --output-debug \
  --output-frames

📖 Citation

If you find this repository useful, please star ⭐ the repository and cite:

@InProceedings{11147940,
  author={Xie-Li, Danny and Fallas-Moya, Fabian},
  booktitle={2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)}, 
  title={PineSORT: A Simple Online Real-Time Tracking Framework for Drone Videos in Agriculture}, 
  year={2025},
  volume={},
  number={},
  pages={65-74},
  keywords={Accuracy;Costs;Scalability;Cameras;Vectors;Robustness;Yield estimation;Object tracking;Drones;Videos;deep learning;agriculture;video;sort},
  doi={10.1109/CVPRW67362.2025.00012}}

Acknowledgements

We thank CeNAT (Centro Nacional de Alta Tecnología) for supporting this research through the CONARE scholarship program (Becas-CONARE), CNCA (Colaboratorio Nacional de Computación Avanzada) for development support, the University of Costa Rica for financial support via project C4612, and the Postgraduate Office of the Costa Rica Institute of Technology (ITCR) for publication assistance.