Towards image-based animal tracking in natural environments using a freely moving camera

Area of Science:

• Ecology
• Ethology
• Computer Vision

Background:

• Image-based animal tracking is crucial for biological and medical research.
• Existing methods often fail to capture unconstrained natural behavior in field settings.

Purpose of the Study:

• To develop a novel visual tracking system for animals in their natural environment.
• To enable the capture of unconstrained animal behavior using freely moving cameras.

Main Methods:

• A global inference method detects animal motion against cluttered backgrounds.
• A key-frame selection scheme and image stitching generate dense animal trajectories on a 2D panorama.
• Minimal constraints on camera orientation and movement are utilized.

Main Results:

• The system accurately extracts per-frame animal positions and overall trajectories.
• It demonstrates effectiveness across various animals, environments, and imaging modalities.
• The method successfully tracks small targets (under 20 pixels) with challenging conditions like poor contrast and occlusion.

Conclusions:

• The developed system provides a flexible and user-friendly approach for obtaining rich data on natural animal behavior.
• It requires only a single uncalibrated camera and no prior training data.
• While rotational drift exists, the system offers valuable qualitative trajectory data within an environmental panorama.