Tail-Wag Stress Detection in Shelter Dogs

ISEF Category: Animal Sciences

Subcategory: Animal Behavior  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Stress shows up in dog body language long before a bark or snap. A tail wag can look like a happy signal, but left-right imbalance may carry a clue you can measure. If you turn that tiny motion into data, you get a project that connects animal behavior, video analysis, and machine learning. That makes a sharp science fair idea.

What Is It?

This project asks whether the way a dog wags its tail can help you spot stress. Tail-wag asymmetry means the tail does not move evenly to both sides. Think of it like a pendulum that swings a little harder one way when something is off. In shelter dogs, that pattern may line up with stress signals in behavior logs.

You would use smartphone video to capture the motion, then train a lightweight CNN, short for convolutional neural network, to learn patterns in the clips. A CNN is a model that finds visual patterns in images and video frames. You are not asking the computer to guess a dog's feelings. You are asking it to learn which motion patterns match logged stress behavior.

Why This Is a Good Topic

This is a good science fair topic because you can test a clear visual pattern against real behavior records. You get a question that is narrow, measurable, and tied to a real problem in shelter welfare. You can learn video labeling, model training, and validation without needing a wet lab. You also get a chance to see whether a simple body-language cue can hold up across different dogs.

Research Questions

• How does tail-wag asymmetry change when a dog moves from a quiet kennel to a busier shelter area?
• What is the effect of dog size or tail type on the CNN's stress prediction accuracy?
• Does the model predict logged stress events better than a simple left-right wag score?
• To what extent do predictions hold up when the model sees dogs it never trained on?
• Which frame rate or clip length gives the cleanest tail asymmetry signal from smartphone video?
• How does the model's accuracy change when you test it on a different shelter or handler team?

Basic Materials

• Smartphone with a stable camera
• Tripod or clamp mount with a fixed angle
• Laptop or desktop computer with Python access
• Spreadsheet software for logs and labels
• Free video annotation tool, such as Label Studio or CVAT
• External storage or a cloud folder for video files
• Printed behavior log template or access to shelter log entries.

Advanced Materials

• High-frame-rate camera or high-end smartphone
• Fixed kennel mount with repeatable camera placement
• GPU-capable workstation or university compute cluster
• Behavior coding software for detailed annotation
• Large monitor or tablet for frame-by-frame review
• Access to shelter behavior logs and kennel metadata
• Low-light or infrared camera for dim kennel areas.

Software & Tools

• Python: Cleans, labels, and analyzes video clips.
• OpenCV: Detects frames, crops tails, and tracks motion.
• TensorFlow/Keras: Trains a lightweight CNN on video features.
• Google Colab: Runs model training without a paid GPU.
• Label Studio: Marks tail positions and behavior tags for training data.

Experiment Steps

1. Define the stress label you will predict and match it to the shelter log fields you can trust.
2. Choose one video setup so camera angle, distance, and lighting stay consistent across dogs.
3. Decide how you will turn each clip into model input, such as cropped frames, motion summaries, or short sequences.
4. Build a baseline rule and a CNN model so you can compare simple scoring with learned predictions.
5. Plan evaluation by dog, not by clip, so the same animal does not appear in both training and test sets.

Common Pitfalls

• Filming from different heights or angles, which changes tail asymmetry even when the dog's mood stays the same.
• Mixing breeds with very different tail shapes, which can make the model key on anatomy instead of stress.
• Using vague shelter notes, which leaves your target label too noisy to train against.
• Putting clips from the same dog in both train and test sets, which inflates accuracy.
• Tracking tail motion during heavy human activity, which teaches the model the kennel scene instead of the dog's state.

What Makes This Competitive

A strong version of this project does more than classify clips. It tests whether tail asymmetry still works after you split data by dog, shelter, and breed type. It also compares the CNN against a simpler baseline, then checks errors against behavior logs, not just raw accuracy. That kind of validation tells you whether the signal is real or just a video artifact.

Project Variations

• Test whether tail-wag asymmetry predicts stress more strongly in intake runs than in kennel rest periods.
• Compare a CNN with a simpler rule-based metric, such as left-right tail angle difference, to see which generalizes better.
• Analyze whether breed size or tail type changes the model's accuracy across dogs.

Learn More

• PubMed: Search review articles on canine stress, body language, and shelter behavior.
• ASPCA Animal Behavior Center: Read free articles on dog stress signals and shelter welfare.
• Shelter Medicine Program at UC Davis: Find free guidance on shelter behavior assessment and animal welfare.
• OpenCV Documentation: Learn the video-processing tools needed for frame extraction and motion tracking.
• TensorFlow Tutorials: Learn how to build and test a lightweight CNN in a free notebook environment.