nmitaeroaiml/buoy-models

🚁 Aerial Search & Rescue: Human & Red Buoy Detection System

A fine-tuned Segment Anything 2 (SAM2) and YOLOv8 aerial search and rescue platform for flood disaster response. It performs real-time aerial human segmentation, hybrid non-circular red buoy detection, human-buoy proximity status tracking, center target reticle positioning, and exports 3 separate mission coordinate breakdown tables for drone payload dropping.

---

🌟 Key Features

1. Fine-Tuned SAM2 Human Segmentation:
   • High-precision SAM2 segmentation powered by fine-tuned weights (best_model.pt, latest_model.pt).
2. YOLOv8 Aerial Human & Flood Detectors:
   • yolov8_human.pt: Fine-tuned aerial human detector for high-altitude drone imagery.
   • yolov8_flood.pt: YOLOv8-OBB model for water surface area ROI filtering.
3. Hybrid Red Buoy Detector:
   • Adaptive Multi-Spectrum HSV & shape-agnostic engine designed for real-world non-circular, oblong, tubular, multi-shaded red/orange buoys under sun glare, shadows, and water spray.
   • Ready-to-load support for fine-tuned yolov8_buoy.pt.
4. Human-Buoy Proximity Status & Target Reticles:
   • Classifies targets into 🚨 Humans WITHOUT Buoy (Priority Drop Targets) vs. ✅ Humans WITH Buoy Nearby (Serviced).
   • High-visibility warning red bounding boxes with center target crosshairs for precision buoy drops.
5. 3 Separate Mission Coordinate Data Breakdown:
   • Interactive pandas DataFrames exportable for drone mission navigation.

---

📁 Repository & Model Weights Structure

.
├── app.py                     # Streamlit Web App & Interactive Dashboard
├── inference.py               # Headless CLI Inference & Evaluation Script
├── train_yolo.py              # YOLOv8 Fine-Tuning Script
├── train_detector.py          # MobileNet Faster R-CNN Detector Training
├── train.py                   # SAM2 Training Pipeline
├── coco_loader.py             # COCO Annotations Dataset Loader
├── preprocess.py              # SAM2 Data Preprocessing
├── requirements.txt           # Python Dependencies
│
├── best_model.pt              # Fine-tuned SAM2 Model Weights (Best)
├── latest_model.pt            # Fine-tuned SAM2 Model Weights (Latest)
├── yolov8_human.pt            # Fine-tuned YOLOv8 Human Detector
├── yolov8_flood.pt            # Fine-tuned YOLOv8-OBB Flood Area Detector
├── sam2_hiera_tiny.pt         # SAM2 Base Tiny Checkpoint
└── sam2_hiera_small.pt        # SAM2 Base Small Checkpoint

---

🚀 Quick Start on Any Machine

1. Clone & Set Up Virtual Environment

git clone https://huggingface.co/YOUR_USERNAME/YOUR_REPO_NAME
cd YOUR_REPO_NAME

python -m venv .venv
# On Windows:
.venv\Scripts\activate
# On Linux/macOS:
source .venv/bin/activate

pip install -r requirements.txt

2. Install SAM2 Package

git clone https://github.com/facebookresearch/segment-anything-2.git sam2_repo
cd sam2_repo
pip install -e .
cd ..

3. Launch Streamlit Dashboard

streamlit run app.py

---

📜 License & Citation

Developed for Search & Rescue flood disaster relief. Powered by PyTorch, Ultralytics YOLOv8, and Meta AI SAM2.