🐾 Wildlife Detection Using Computer Vision

An AI-powered learning-focused project built to understand how real-time computer vision systems behave under practical constraints such as CPU-only inference, frame drops, and visualization stability.

This project uses a trained YOLOv8 model to detect wildlife animals from a live screen capture or camera feed. Detected animals are visualized in real time using bounding boxes and confidence scores.

⚠️ Note: This project is built primarily for learning and experimentation, with emphasis on:

real-time detection behavior

CPU-based inference limitations

bounding box stabilization

confidence filtering —not on achieving perfect accuracy or production deployment.

---

🚀 Key Features

Real-time wildlife detection using a trained YOLOv8 (best.pt) model

Live visualization with bounding boxes and confidence scores

CPU-friendly inference with configurable thresholds

Bounding box smoothing to reduce flickering

Modular design that can be extended to cameras, videos, or backend systems

---

🧠 Tech Stack

AI / Computer Vision :- YOLOv8 (Ultralytics),OpenCV,NumPy

Languages :- Python

(Backend intentionally excluded to preserve real-time performance)

---

🏗️ Project Structure

Wildlife-Detection/ ├── model/ # YOLO Model and Scripts │ ├── best.pt # Trained YOLOv8 model (use Git LFS or Drive link) │ └── test_model.py # Detection script ├── README.md ├── .gitignore └── .env

---

🧩 How It Works (High-Level)

Capture frames from a live screen or camera feed

Resize frames for efficient CPU-based inference

Run YOLOv8 object detection on each frame

Filter detections using confidence thresholds

Stabilize bounding boxes to reduce jitter

Visualize detections in real time

---

🧩 Setup Instructions

🐍 Run YOLO Model (Python)

pip install ultralytics pyautogui opencv-python numpy
python test_model.py
Ensure `best.pt` is placed inside the `model/` directory as shown in the project structure.

---

📦 Model File

Model weights (best.pt) can be downloaded from: 🔗 https://drive.google.com/file/d/1HKvEeGQ4RfjMRXvr3Psdlh6arvQiwjZv/view?usp=sharing

---

📈 Learnings & Observations

Real-time CV systems behave very differently than offline demos

CPU-only inference requires careful tuning of resolution and thresholds

Confidence scores strongly affect false positives

Bounding box smoothing significantly improves visual stability

Temporary detection loss is normal in real-world input streams

---

🔮 Future Improvements

Improve class separation and reduce misclassification

Experiment with different confidence and IoU thresholds

Optimize further for low-resource devices

Optional backend integration for analytics (async / non-blocking)

---

🧑‍💻 Author

Divyansh Kande

Full Stack Java + AI Developer

🔗 GitHub: https://github.com/divyanshkande

⭐ If you find this project interesting or useful, feel free to explore, fork, or share feedback.

---