Hybrid Deep Learning-Based Human Activity Recognition Enhanced by YOLOv8 and Data Augmentation on MHealth and WISDM Datasets

Objectives: To propose an advanced Human Activity Recognition Framework to classify complex human activities with high accuracy and computational efficiency by integrating hybrid deep learning and optimization techniques. The model helps to build a robust system capable of maintaining high performance across diverse user behaviours and heterogeneous sensor modalities. Methods: Hybrid deep learning YOLOv8 extracts rapid spatial features from image and video frames. Vision Transformer (ViT) enhances the classification depth and captures long-range dependencies and global context. To ensure optimal tuning, particle swarm optimization (PSO) is applied for hyperparameter optimization, which includes batch size, learning rate, and drop ratio. Synthetic data augmentation, such as cropping, rotation, flipping, etc., is done using GAN-based generation, which helps to enhance generalization. The YOLOv8-ViT with PSO model extracts data from two benchmark datasets, WISDM and MHealth, comprising physical activities, multi-dimensional time series data, and annotated sensor readings. The proposed model is evaluated using MATLAB and Python, and the results are compared with existing models such as HLA, SMO-DNN, and AMC-CNN. Findings: The proposed model yields promising results with 95.4% accuracy, 96.6% sensitivity, 96.8% specificity, 0.92 AUC, 93.2% MCC, and 94.6% F1-Score, which is substantially outperforming the existing computational models for HAR with different datasets. Novelty: This research presents a multi-layered architecture combining YOLOv8’s speed, contextual power of ViT, PSO’s adaptive tuning, and GAN-based data augmentation, which offers scalable, real-time, and high precision HAR models for future healthcare and innovative sensing applications. Keywords: Human Activity Recognition (HAR), YOLOv8, Deep Learning, Vision Transformer, Particle Swarm Optimization, Data Augmentation