Non-Invasive Pipeline for Detecting Normal and Unusual Human Activities Using 2D Skeleton-Based Features and Categorical Boosting Models

Abstract This paper introduces a non-invasive, pose-based pipeline for classifying normal (eating snacks, sitting quietly, walking, using phone) and unusual (head banging, throwing things, attacking, biting) behaviors using 2D skeleton data from five subjects. Seventeen keypoints are first normalized by centering on the hip midpoint and scaling by the average shoulder-to-shoulder and hip-to-hip distances to ensure consistent scale and orientation across subjects. Time-series features are then extracted with TSFEL and customized features, after which a preliminary Categorical Boosting (CatBoost) model ranks feature importances, and the top 500 features are retained. A CatBoostClassifier trained under Leave-One-Subject-Out cross-validation achieves an average accuracy of 76.29% with a per-class Macro F1 score of 75.06%. Main contributions include an end-to-end framework for detecting both normal and unusual behaviors from 2D skeleton sequences, a comprehensive analysis of discriminative time-series features that drive classification performance, and empirical evidence demonstrating strong cross-subject generalization under Leave-One-Subject-Out evaluation.