EFFICIENT POSE-BASED HUMAN ACTIVITY RECOGNITION THROUGH SKELETAL FEATURE ENGINEERING AND DIMENSIONALITY REDUCTION

Abstract

Human Activity Recognition (HAR) plays a vital role in applications such as fitness monitoring, rehabilitation, surveillance, and intelligent human-computer interaction sys tems. While recent deep learning approaches based on RGB video data have achieved high accuracy, they are computationally expensive, require large-scale datasets, and are sensi tive to variations in background and illumination. To address these limitations, this thesis explores lightweight and efficient pose-based approaches for human activity recognition using skeletal representations. This work presents a comprehensive study of pose-based HAR through two complementary approaches. A comparative analysis of classical ma chine learning models, including Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), Logistic Regression (LR), and XGBoost, is conducted us ing statistical features extracted from MediaPipe-based skeletal keypoints. Experiments across benchmark datasets such as UCF101, PennAction, and HAA500 demonstrate that traditional machine learning models can achieve competitive performance with signifi cantly lower computational requirements and a lightweight deep learning framework is proposed that leverages engineered skeletal features incorporating spatial joint coordi nates, temporal motion dynamics, joint angles, and limb distances. Temporal modeling is performed using overlapping sliding windows, followed by dimensionality reduction using Principal Component Analysis (PCA) to retain 95% variance while significantly re ducing feature dimensionality. A feed-forward neural network (FFNN) is employed for classification. The proposed method achieves high recognition performance, with up to 96% window-level accuracy and 98% video-level accuracy on UCF101, and comparable iv results on PennAction. The results demonstrate that well-designed skeletal feature representations combined with efficient machine learning and dimensionality reduction techniques can serve as a pow erful alternative to computationally intensive deep learning models. This thesis high lights the potential of lightweight, pose-based HAR systems for real-time and resource constrained environments, while maintaining strong recognition performance.