A Novel Activity Pattern Recognition via Convolutional Neural Networks and Advanced Skeleton Models
Tanvir Fatima Naik Bukht, Naif S. Alshassabi, Haifa F. Alhasson, Bayan Ibrahimm Alabdullah, Ahmad Jalal · Traitement du signal · 2025
Human Activity Recognition (HAR) is crucial to intelligent smart home systems.In this research, we propose a novel skeleton-based method for recognizing human activities accurately.Gamma correction is applied as a preprocessing step to improve image quality.Then, we use a robust combination of Multiple Object Tracking (MOT) and graph-based segmentation techniques to extract precise human silhouettes from video sequences.This research also introduces a novel innovation in developing a 23-joint skeleton model that accurately identifies and tracks key body joints.A comprehensive set of features extracted from this skeleton data is derived, including relative joint angles, joint proximity measures, joint stability, and full body features, which are extracted using BRIEF, LATCH, and MSER.A fuzzy optimization technique is employed to find the most discriminative features to optimize feature selection.Finally, a Convolutional Neural Networks (CNN) classifier is trained on the optimized features to classify human activities accurately.Experimental results demonstrate the effectiveness of our approach, with ShakeFive2 achieving an 88% accuracy rate and BIT-Interaction achieving 94% on a benchmark dataset.This work contributes to advancing human activity understanding in various domains, such as surveillance, human-behavior interaction, healthcare, sports, and social robotics.