Hybrid Human Activity Recognition: Integrating Traditional Feature Engineering with Deep Learning Approach

Abstract

Human Activity Recognition (HAR) is a vital research area with applications in healthcare, security, and intelligent environments. This paper presents a hybrid framework that combines traditional feature engineering with deep learning to enhance HAR performance. It leverages the Histogram of Oriented Gradients (HoG) for spatial feature extraction and Support Vector Machines (SVM) for structured classification. Additionally, Vision Transformers (ViT) and ResNet architectures are integrated to improve accuracy: ViT captures global dependencies through attention mechanisms, while ResNet enhances deep feature learning through skip connections. Experimental results demonstrate that this approach balances computational efficiency, interpretability, and high accuracy on large datasets.