Related Experiment Video
Updated: Jun 3, 2025
11:04
Biomolecular Detection employing the Interferometric Reflectance Imaging Sensor IRIS
Published on: May 3, 2011
14.6K
A Comparison Study of Person Identification Using IR Array Sensors and LiDAR
Kai Liu1, Mondher Bouazizi2, Zelin Xing1
1Graduate School of Science and Technology, Keio University, Yokohama 223-8522, Japan.
Sensors (Basel, Switzerland)
|January 11, 2025
Summary
Thermal imaging offers the most robust person identification across varying resolutions, outperforming RGB and depth data. This modality demonstrates superior generalizability by focusing on subject features, crucial for reliable security and surveillance systems.
Area of Science:
- Computer Vision
- Machine Learning
- Biometrics
Background:
- Person identification is vital for security and surveillance.
- Existing systems require robust performance under diverse conditions.
- Evaluating deep learning models across multiple data modalities is essential.
Purpose of the Study:
- To evaluate the effectiveness of Vision Transformer (ViT) and ResNet34 models for person identification.
- To compare performance across RGB, thermal, and depth modalities under varying resolutions.
- To identify the most reliable modality for robust person identification.
Main Methods:
- Utilized ViT and ResNet34 models for person identification.
- Tested models on RGB, thermal, and depth datasets captured by infrared and LiDAR sensors.
- Employed YOLO-based cropping for subject isolation and analyzed performance across resolutions from 16x12 to 640x480.
Main Results:
- High identification performance across modalities at high resolution (640x480): RGB (100.0%), depth (99.54%), thermal (97.93%).
- Thermal images demonstrated superior robustness and generalizability at low resolutions, focusing on subject features.
- RGB performance degraded at low resolutions due to background reliance; depth data suffered from artifacts and scattered attention.
Conclusions:
- Thermal imaging is the most reliable modality for person identification, especially in low-resolution scenarios.
- Modality selection is critical for designing robust person identification systems.
- Future research should focus on multi-modal integration and advanced architectures for enhanced adaptability.
