A Machine Learning Framework for Automated Computational Ethology Using Markerless Pose Estimation

Quantitative behavioral analysis is fundamental to ethological research, yet automated approaches remain limited by the gap between pose estimation and meaningful behavioral classification. Most existing methods focus on either pose detection or behavior recognition in isolation, lacking integrated frameworks for comprehensive behavioral analysis. We present an end-to-end framework that bridges markerless pose estimation with machine learning classification for automated behavioral analysis. Our framework integrates SLEAP pose estimation, systematic feature engineering, multiple machine learning algorithms, and robust validation strategies into a unified pipeline. We demonstrate the framework on Drosophila larvae videos, automatically classifying three behavioral states (feeding, sleeping, crawling) from pose trajectories. We evaluate five machine learning models across three validation strategies and engineer twelve position-invariant features from four anatomical landmarks. The framework provides computational ethology researchers with practical tools for pose-based behavioral classification, comprehensive model evaluation, and deployment guidance for real-world applications.