# Non-contact human identification through radar signals using convolutional neural networks across multiple physiological scenarios

## TL;DR

This paper explores using radar signals and neural networks to identify people non-contactly, even when their physiological state changes.

## Contribution

The first study to explore radar-based human identification across multiple physiological scenarios using deep learning.

## Key findings

- A CNN-based method achieved 97.70% accuracy in identifying subjects in a resting state.
- The system maintained high accuracy (98.6%) in high-confidence predictions across different physiological scenarios.
- Radar-based identification can rival contact-based biometric methods like ECG or PPG.

## Abstract

In recent years, contactless identification methods have gained prominence in enhancing security and user convenience. Radar-based identification is emerging as a promising solution due to its ability to perform non-intrusive, seamless, and hygienic identification without physical contact or reliance on optical sensors. However, being a relatively new technology, research in this domain remains limited. This study investigates the feasibility of secure subject identification using heart dynamics acquired through a continuous wave radar. Unlike previous studies, our work explores identification across multiple physiological scenarios, representing, to the best of our knowledge, the first such exploration.

We propose and compare two identification methods in a controlled Resting scenario: a traditional machine learning pipeline and a deep learning-based approach. The latter consists of using a Convolutional Neural Network (CNN) to extract features from scalograms, followed by a Support Vector Classifier (SVC) for final classification. We further assess the generalizability of the system in multiple scenarios, evaluating performance both when the physiological state is known and when it is not.

In the Resting scenario, the deep learning-based method outperformed the traditional pipeline, achieving 97.70% accuracy. When extending the identification task to various physiological scenarios, 82% of predictions exceeded scenario-specific confidence thresholds, achieving 98.6% accuracy within this high-confidence subset.

Our findings suggest that radar-based identification systems can match the performance of established biometric methods such as electrocardiography (ECG) or photoplethysmography (PPG), while offering the additional benefit of being contactless. This demonstrates the potential of radar heart signal analysis as a reliable and practical solution for secure human identification across diverse conditions.