Study Introduces Tensor-Based Model for Enhanced Human Activity Recognition

Global: Study Introduces Tensor-Based Model for Enhanced Human Activity Recognition

A research team has unveiled a low‑cost, automated framework for human activity recognition (HAR) that leverages wearable inertial sensors and a novel machine‑learning approach. The system, detailed in a recent arXiv preprint, targets remote health monitoring for elderly and vulnerable populations by classifying motions such as walking, stair navigation, sitting, standing, and lying. Using accelerometer and gyroscope data, the authors evaluated several conventional classifiers before introducing a Support Tensor Machine (STM) that achieved the highest reported accuracies.

Data Acquisition and Sensor Setup

Participants performed six distinct activities while equipped with wearable devices that recorded three‑axis acceleration and angular velocity. The collected dataset captured the spatio‑temporal dynamics of each motion, providing a basis for training and testing various classification models.

Benchmark Classifier Results

Four classical algorithms—Logistic Regression, Random Forest, Support Vector Machine (SVM), and k‑Nearest Neighbors (k‑NN)—were trained on the sensor data. SVM attained an accuracy of 93.33 percent, whereas Logistic Regression, Random Forest, and k‑NN each reached 91.11 percent, establishing a performance baseline for the proposed approach.

Support Tensor Machine Performance

The introduced Support Tensor Machine leveraged tensor representations to preserve the inherent motion dynamics across time and space. In testing, STM achieved a classification accuracy of 96.67 percent, surpassing the conventional models. Moreover, cross‑validation yielded an accuracy of 98.50 percent, indicating strong generalization potential.

Implications for Remote Healthcare

By delivering reliable activity monitoring without extensive infrastructure, the framework could support home‑based therapeutic programs, such as yoga or physiotherapy, particularly in low‑resource or rural settings. Accurate detection of movement patterns may help caregivers identify non‑adherence or safety concerns promptly.

Future Directions

The authors suggest extending the system to larger, more diverse populations and exploring real‑time deployment on edge devices. Integrating additional sensor modalities and refining tensor‑based algorithms could further improve robustness across varied activities and environments.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.