This project leverages deep learning and image slicing techniques to detect and count Sandhill Cranes in aerial thermal imagery. It combines object detection models (YOLOv5) with efficient inference strategies (SAHI) and is being extended into a fully deployable app with MLOps best practices.
📌 Ongoing collaboration with wildlife conservation efforts to monitor crane populations in protected habitats.
- Detect Sandhill Cranes in aerial thermal images using deep learning
- Improve inference accuracy using Sliced Aided Hyper Inference (SAHI)
- Build a user-facing web app where users can upload images and receive bird counts with bounding boxes
- Refactor training/inference pipelines for reproducibility and deployment
- Deploy the app using AWS + Docker + Streamlit
- YOLOv5: Object detection architecture (via Ultralytics)
- SAHI: Efficient inference on large images through slicing
- PyTorch: Model training & experimentation
- OpenCV: Image preprocessing
- FastAPI: Lightweight web framework for serving predictions
- Docker: Containerized deployment
- AWS S3 + Lambda (planned): Hosting and model serving
- GitHub Actions + DVC: CI/CD and data/model versioning
- Next.js (v15): React-based framework with App Router
- TypeScript: Strongly typed JavaScript for scalable development
- Tailwind CSS: Utility-first CSS for rapid UI styling
- 🖼️ Handles large aerial images with small, sparse bird targets
- 📦 Modular pipeline for training, evaluation, and inference
- 📸 Returns output image with bounding boxes and crane count
- 🌐 Web app in development to allow non-technical users to interact with the model