IMU_HardBiometrics

Introduction

This project provides the code for the experiments of hard biometrics based on IMU gait sequences. It serves as an important part of the author's master thesis (30ECTS) in ETH Zurich.

Summary

This project mainly comprises five parts:

1. Segmenting the original IMU gait signals into IMU sequences of equal length. (csv format)
2. Transforming the IMU sequences into hand-crafted features. (csv format)
3. Transforming the IMU sequences into autocorrelation features. (csv format)
4. Evaluating the performance of all included algorithms for hard biometrics.
5. Evaluating the underlying bias of hard biometrics models.

Package Requirements

• matlabengine==23.2
• numpy==1.26.4
• pandas==2.2.2
• PyYAML==6.0.1
• scikit_learn==1.4.1.post1
• scipy==1.13.1
• statsmodels==0.14.1
• torch==2.2.1+cu118
• tqdm==4.66.2
• tsai==0.3.9

Configuration

The experiments were runned on the author's personal laptop. The configurations are provided as the reference:

• CPU: AMD Ryzen 9 5900HX with Radeon Graphics
• GPU: NVIDIA GeForce RTX 3080 Laptop GPU
• CUDA Version: 11.7
• Operating System: Microsoft Windows 11 (version-10.0.22631)

Code Structure

IMU_HardBiometrics
├─config
├─data
│  └─OU_ISIR_Inertial_Sensor
│      ├─manual_IMUZCenter
│      │  ├─data_segmented
│      │  ├─feature_autocorr
│      │  └─feature_manual
│      ├─manual_IMUZLeft
│      │  ├─data_segmented
│      │  ├─feature_autocorr
│      │  └─feature_manual
│      └─manual_IMUZRight
│          ├─data_segmented
│          ├─feature_autocorr
│          └─feature_manual
├─main
├─result
│  └─OU_ISIR_Inertial_Sensor
│      ├─manual_IMUZCenter
│      ├─manual_IMUZLeft
│      └─manual_IMUZRight
└─util

Datasets

1. OU-ISIR Inertial Sensor Dataset: http://www.am.sanken.osaka-u.ac.jp/BiometricDB/InertialGait.html
2. OU-ISIR Similar Action Inertial Dataset: http://www.am.sanken.osaka-u.ac.jp/BiometricDB/SimilarActionsInertialDB.html

Usage

First, activate the local environment and then set the folder containing this README file as the current folder.
For Windows, execute: python (...).py
For Linux, execute: python3 (...).py

1. Segment the original IMU gait signals into IMU sequences of equal length: python "./main/main_DataSegmentation.py"
2. Transform segmented IMU signals into hand-crafted features: python "./main/main_IMU2ManualFeature.py"
3. Transform segmented IMU signals into autocorrelation features: python "./main/main_IMU2AutoCorrFeature.py"
4. Evaluate the performance of Combined Feature + KNN for hard biometrics: python "./main/main_HumanAuthentication_combined.py"
5. Evaluate the performance of HYDRA for hard biometrics: python "./main/main_HumanAuthentication_hydra.py"
6. Evaluate the performance of InceptionTime for hard biometrics: python "./main/main_HumanAuthentication_inception.py"
7. Generate the recognition result of Combined Feature + KNN for bias evaluation: python "./main/main_HumanAuthentication_combined_bias.py"
8. Generate the recognition result of HYDRA for bias evaluation: python "./main/main_HumanAuthentication_hydra_bias.py"
9. Generate the recognition result of InceptionTime for bias evaluation: python "./main/main_HumanAuthentication_inception_bias.py"
10. Quantitatively evaluate the False Match Rate (FMR) and False Non-Match Rate (FNMR): python "./main/main_PredictiveParity.py"
11. Correlation analysis of age for misidentification pairs: python "./main/main_AnalysisAge.py"
12. Quantify misidentification pairs of same/different gender: python "./main/main_AnalysisGender.py"

Example Results

User Recognition (Training: walk-1. Testing: walk2.)

Baseline Method	IMUZ-Left	IMUZ-Center	IMUZ-Right
Combined Feature + KNN	80.66%	79.72%	75.74%
Original Signal + InceptionTime	81.02%	84.86%	78.92%
HYDRA	95.26%	94.85%	92.79%

Contact

If you have any questions, please feel free to contact me through email (shuoli199909@outlook.com)!

Authors and acknowledgment

This master thesis was supervised by Dr. Mohamed Elgendi (ETH Zurich), Prof. Dr. Carlo Menon (ETH Zurich), and Prof. Dr. Rosa Chan (City University of Hong Kong). The code was developed by Shuo Li. Also thank to all my colleagues and providers of datasets for the continuous help!

License - MIT License.