PriMAT: Robust multi-animal tracking of primates in the wild

Area of Science:

• Computer Vision
• Ethology
• Machine Learning

Background:

• Automated animal behavior studies rely on video analysis.
• Current deep learning tracking methods excel in lab settings but struggle with wild environments due to variability and occlusion.
• Multi-animal tracking in the wild requires robust solutions for complex scenarios.

Purpose of the Study:

• To develop a flexible and robust approach for detecting and tracking nonhuman primates in the wild.
• To enable individual identification of animals for enhanced behavioral analysis.
• To provide a tool for studying animal behavior in natural habitats.

Main Methods:

• Proposed PriMAT, a one-stage deep learning model using bounding boxes instead of keypoints for detection and tracking.
• Incorporated a classification branch for individual identification.
• Trained and evaluated the model on wild Assamese macaques and redfronted lemurs, with transfer learning applied to other primate species.

Main Results:

• PriMAT achieved robust tracking results with limited labeled data (few hundred frames).
• The lemur identification model demonstrated 84% accuracy in predicting individual identities.
• The approach showed successful transferability to various primate species and settings.

Conclusions:

• PriMAT offers a promising solution for accurate wild animal tracking and identification.
• The bounding box approach simplifies data annotation and enhances robustness.
• Publicly available code and data facilitate further research in animal behavior analysis.