Zoo Behavior Scoring With CNNs

ISEF Category: Animal Sciences

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

The Hook

A pacing tiger can tell you more than a worksheet ever could. Zoo teams watch for stereotypic behaviors, repeated movements like pacing or swaying, because they can point to stress or poor welfare. Your project can test whether a CNN can score those behaviors from video clips. That gives you a real mix of animal welfare and computer vision.

What Is It?

A welfare-scoring rubric turns animal behavior into a repeatable score. In this project, the rubric looks for stereotypic behaviors, repeated actions like pacing or swaying that may show stress, boredom, or limited space. Think of it like a warning light that keeps blinking instead of a one-time event.

A CNN, or convolutional neural network, is a kind of machine learning model that finds patterns in images and video frames. Instead of checking every clip by hand, the model learns to notice movement, body position, and scene cues. Your goal is to see whether that score stays reliable when the animal, camera angle, or lighting changes.

Why This Is a Good Topic

This topic works well because you can measure it with public video, clear labels, and standard model metrics. It connects to a real animal welfare problem, since zoos and researchers need fast ways to spot repeated stress behaviors. You can also learn how to clean data, label behavior, test model accuracy, and check whether a model works beyond the clips it trained on.

Research Questions

• How does video resolution affect a CNN's ability to detect pacing in zoo animal clips?
• What is the effect of clip length on stereotypic behavior detection accuracy?
• Does training on one species and testing on another reduce welfare score accuracy?
• To what extent do lighting changes alter false positive rates for swaying detection?
• Which frame sampling rate gives the best balance of speed and accuracy for video scoring?
• How does adding more labeled examples of rare behaviors change recall?

Basic Materials

• Laptop with at least 8 GB RAM.
• Internet access to download public zoo-animal video clips.
• Spreadsheet for manual labels and notes.
• External storage or cloud folder for video files.
• Notebook for writing labeling rules and error checks.
• Headphones for checking clips with audio cues, if present.

Advanced Materials

• GPU workstation or university server with a CUDA-capable graphics card.
• Curated archive of zoo video with species and site metadata.
• Secure high-capacity storage for video files and labels.
• Reference ethogram or welfare rubric from a zoo advisor.
• Annotation workstation for frame-by-frame review.
• Statistical software for mixed-effects analysis.

Software & Tools

• Python: Cleans labels, trains models, and runs evaluation scripts.
• Google Colab: Provides a free notebook environment if your laptop is slow.
• Label Studio: Helps you tag pacing and swaying in video clips.
• OpenCV: Reads video frames and extracts motion features.
• scikit-learn: Scores precision, recall, F1, and confusion patterns.

Experiment Steps

1. Define one target behavior and write a short labeling rule for what counts as a positive clip.
2. Split your video set by animal or zoo first, then decide which clips belong in train, validation, and test.
3. Choose the output you want to score, such as clip-level welfare risk or frame-level behavior presence.
4. Plan a baseline model and a comparison model so you can tell whether your design choices help.
5. Predefine the metrics you will report, including precision, recall, F1, and confusion by behavior class.
6. Decide which outside factors, like species or camera angle, you will hold constant or test on purpose.

Common Pitfalls

• Mixing clips from the same animal across train and test sets, which makes accuracy look higher than it really is.
• Labeling normal grooming or stretching as stereotypy, which blurs the class boundary.
• Letting enclosure bars, feeders, or camera zoom become shortcuts the model learns instead of the behavior.
• Using too few positive examples of pacing or swaying, which makes recall swing wildly.
• Reporting only overall accuracy, which hides poor performance on the rare welfare problem you care about.

What Makes This Competitive

A stronger version of this project tests whether one rubric holds up across species, enclosures, and camera styles instead of only one zoo and one animal. You can compare a CNN against human labels, then report precision, recall, and confusion by behavior type. A tougher project also checks whether the model stays accurate when you hold out an entire animal or zoo site. That kind of design shows you understand generalization, not just training accuracy.

Project Variations

• Compare pacing detection in big cats, primates, and hoofstock to see whether species changes the model's error pattern.
• Test whether a frame-based CNN or a short-clip model does a better job of spotting swaying in public zoo footage.
• Swap video clips for still frames and compare how much motion information the welfare score depends on.

Learn More

• PubMed: Search review articles on stereotypic behavior, zoo animal welfare, and behavior classification.
• PubMed Central: Find free full-text papers on animal behavior recognition and zoo welfare.
• Journal of Zoo and Aquarium Research: Read peer-reviewed studies on zoo animal behavior and welfare through the journal site.
• arXiv: Look for preprints on video action recognition and animal behavior detection.
• MIT OpenCourseWare: Review free lectures on machine learning and computer vision to understand model training and evaluation.