Important
This is the finalized main branch of the Project. (Submitted for the PU)
This repository contains a complete MATLAB-based pipeline for biometric user identification and verification using Accelerometer and Gyroscope data. The system processes raw motion signals, extracts high-dimensional features, and utilizes Feed-Forward Neural Networks (FNN) to authenticate users.
Warning
Note on Reproducibility: To ensure consistent results, a global random number generator seed (rng) is implemented in the training scripts. Minor variations in EER may occur if the seed is modified or stochastic initialization is reset.
The system transforms raw time-series sensor data into a biometric profile. It evaluates performance using two distinct protocols:
- Cross-Day : Training on data from one day and testing on another (evaluates temporal generalization).
- Combined-Day: Randomly shuffling all available data for training and testing.
- MATLAB
- Required Toolboxes:
- Signal Processing Toolbox
- Statistics and Machine Learning Toolbox
- Deep Learning Toolbox
- Clone this repository.
Execute the scripts in the following order:
| Step | Script Name | Description |
|---|---|---|
| 1 | a_01_Data_Preprocessing.m |
Imports CSVs, performs Z-score outlier removal, and applies smoothing filters. |
| 2 | a_02_Feature_Engineering_Baseline.m |
Segments signals into 3s windows and extracts 180+ Time/Frequency features. |
| 3 | a_03_FNN_Feature_Optimization.m |
Reduces features via ANOVA ranking and SFS optimization. |
| 4 | a_04_FNN_CrossDay_Model_Evaluation.m |
Evaluates system generalization using the Cross-Day protocol (FD → MD). |
| 5 | a_04_FNN_CombinedDay_Model_Evaluations.m |
Evaluates the system using the Combined-Day (Randomized) protocol. |
| 6 | a_05_FNN_Optimized_Eval.m |
Final evaluation using optimized feature sets and cross-day verification. |
- Filtering: Median filter for spike removal and Moving Average filter for noise reduction.
- Normalization: Z-score normalization to ensure feature scaling consistency.
- Segmentation: 3-second non-overlapping windows (31Hz sampling frequency).
- Initial Ranking: ANOVA F-Statistic identifies high-variance features between users.
- Optimization: Sequential Feature Selection (SFS) identifies the optimal subset (20 features) for cross-day stability.
- Type: Feed-Forward Multi-Layer Perceptron (MLP).
- Hidden Layer: Optimized at 12 hidden nodes.
- Output: Softmax layer providing probability scores for One-vs-Rest (OvR) verification.
Performance metrics derived from the optimized evaluation logs.
| Protocol | Best Model | Features | Equal Error Rate (EER) |
|---|---|---|---|
| Cross-Day (Generalization) | ACC Total | 20 | 1.24% |
| Combined-Day (Random Split) | Full Combined | 20 | 0.40% |
- Sensor Robustness: Accelerometer-based models (ACC) significantly outperformed Gyroscope (GYR) models, indicating foot-strike impact provides more stable biometric signatures than torso rotation.
- Feature Optimization: Implementing SFS and ANOVA reduced the feature space from 98 to 20 dimensions while improving Cross-Day EER by over 5% compared to non-optimized baselines.
- Temporal Stability: While the model achieved near-perfect results on same-day data (0.40% EER), the Cross-Day evaluation (1.24% EER) confirmed the system's ability to handle natural gait variance over time.
- Classifier: Feedforward Neural Network (FNN-MLP) with 12 Hidden Nodes.
-
Data Processing: 3-second non-overlapping windows (
$F_s = 31$ Hz). - Validation: One-vs-Rest (OvR) verification protocol used to determine true False Acceptance Rates (FAR) and False Rejection Rates (FRR).
This project is developed for research and educational purposes in the field of Biometrics and Human Activity Recognition. All rights reserved.