DeepF-SVM: A new hybrid deep learning model for enhanced sensor-based human activity recognition

Download PDF
Open Webpage

Imene Charabi, M'hamed Bilal Abidine, Belkacem Fergani, Mourad Oussalah

Published: 2025, Last Modified: 06 Nov 2025Clust. Comput. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Human Activity Recognition (HAR) has long been a research hotspot in the pattern recognition field due to its extensive applications across various domains. The core idea of HAR is to train machines to identify human physical activities using data recorded by various sensor modalities, which is particularly useful in areas such as e-health, where fall detection and remote patient health monitoring are of paramount importance. Traditional machine learning algorithms, such as Support Vector Machines (SVM), have demonstrated strong performance in HAR state-of-the-art literature; however, they rely on manual feature extraction, which is time-consuming and requires domain expertise. In contrast, recent advancements have established Convolutional Neural Networks (CNNs) as powerful tools that automatically extract optimal features directly from raw data, eliminating the need for manual intervention. In this paper, we introduce a hybrid model called DeepF-SVM to enhance the performance of CNNs and address the reliance of SVM on domain expertise. First, a one-dimensional CNN with three convolutional layers is trained on raw sensor data to extract deep features (DeepF). Then, an SVM classifier with an RBF kernel replaces the final dense layer of the CNN, taking the DeepF from the preceding layer as input for activity classification. Experiments are conducted on three publicly available datasets–UCI HAR, UniMiB SHAR, and PAMAP2–to evaluate the performance of the proposed approach. The DeepF-SVM model achieved accuracy scores of 96.44%, 93.57%, and 98.48% on the above three datasets, respectively, with inference times of 0.3175s for UCI HAR, 1.1168s for UniMiB SHAR, and 0.3672s for PAMAP2. The results demonstrate that the developed DeepF-SVM model outperformed both standalone CNN and standalone SVM models, confirming its high effectiveness and potential prospects in HAR tasks.