Human Activity Recognition through Skeleton-Based Motion Analysis Using YOLOv8 and Graph Convolutional Networks

Abstract

Human Activity Recognition has become an important research domain in developing intelligent systems for sectors such as healthcare, behavioral analytics, and surveillance monitoring. Traditional vision-based HAR approaches have limitations in terms of subject variability, occlusion, and background clutter. To address this, a novel skeleton-based motion analysis model is proposed to enhance the precision and temporal understanding of human motions by combining real-time keypoint extraction with graph-structured spatial-temporal learning. The proposed YOLOv8 + Graph Temporal Convolution for Human Activity Recognition (YGTC-HAR) consists of four essential stages, including: (1) YOLOv8-Pose to detect human figures in real-time, and (2) Graph Convolutional Network (GCN) is used to transform the joint coordinates into a graph representation graph representation. (3) The Temporal Convolutional Network (TCN) is designed to learn the sequential motion dynamics and time-dependent characteristics of human activities. Additionally, Genetic Algorithm (GA) and Bayesian Optimization (BO) are adopted to fine-tune hyperparameters, including learning rate, dropout ratio, and convolutional filters. MHealth and WISDM datasets are utilized in this research to enable comprehensive testing across static and dynamic movements. The proposed YGTC-HAR is implemented using Python (with TensorFlow and PyTorch) for deep learning, and MATLAB R2023b is used for signal processing, graphical visualization, and performance validation. The proposed work is compared against existing HLA, SMO-DNN, AMC-CNN, and YOLOv8-ViT models. The model achieves 97.6% accuracy, 98.4% sensitivity, 97.8% specificity, 97.2% F1-score, 96.4% MCC, and an AUC of 0.96, which outperforms the existing models by over 4.3%. The proposed YGTC-HAR serves as a single end-to-end HAR framework that delivers superior generalization, real-time performance, and reliability for HCIA (Human-Centered Intelligent Applications). The novelty of the model lies in the combination of YOLOv8-driven skeleton extraction, GCN-based spatial modeling, TCN-driven temporal learning, and adaptive optimization.