Introduction
Studying how primates navigate and remember their environment has historically relied on manual observation. Researchers typically recorded visits to specific "goalboxes" using pen and paper, a method prone to human error and incapable of capturing continuous path data, travel speed, or behavioral hesitations. While automated tracking is common in controlled laboratory settings, outdoor environments pose significant challenges for computer vision due to shifting shadows, weather changes, and reflections.
This paper introduces a semi-automated, two-camera system designed to bridge this gap, providing a robust way to extract 3D spatial data in complex real-world settings.
Methodology
The tracking system utilizes a pair of calibrated digital cameras to monitor a 4.74m x 4.74m outdoor area at the Yerkes National Primate Research Center. The technical pipeline involves several critical stages:
- Adaptive Background Subtraction: Pixels are modeled using a mixture of Gaussian distributions to distinguish fast-moving subjects from slow environmental changes like moving shadows.
- LUV Color Space Conversion: By separating luminance from chromaticity, the system maintains tracking stability under varying outdoor lighting conditions.
- 3D Triangulation: The system calculates 3D coordinates five times per second based on the 2D center of the subject in both camera feeds.
- Interactive Correction: A "human-in-the-loop" interface allows researchers to manually "nudge" the tracker if it loses the subject, allowing a 5-minute trial to be processed in just 1–2 minutes.
Results and Key Insights
The system's accuracy was validated using a human walking a known path, showing an average absolute difference of only 2.03% (± 7.14%) in path length estimation.
A significant finding was the relationship between speed and memory. Monkeys moved faster when heading toward a goalbox they remembered contained food and slower when approaching unknown or incorrect locations. This suggests that instantaneous speed can serve as a quantitative proxy for an animal's confidence. Furthermore, combining distance-to-goal with speed thresholds (e.g., moving slower than 0.05 m/sec) reduced false positives in automated visit detection from 65 to 46.
Conclusion
The developed system offers a significant advancement over manual observation, providing high-fidelity spatial data with minimal human intervention. While stationary subjects still pose a challenge for the adaptive algorithm, future work involving machine learning could further automate behavior recognition. This technology is currently being applied to investigate sex differences in how rhesus monkeys navigate space.
Summary updated