Gait-Based User Authentication

MATLAB machine learning system for user identification using walking patterns from wearable sensors.

Overview

Identifies users based on unique gait patterns using accelerometer and gyroscope data. Extracts 74 features and trains 3 ML models (KNN, SVM, Neural Network) for authentication.

Features

• 6-axis IMU sensor processing (accelerometer + gyroscope)
• 74 time/frequency domain features
• Multiple classifiers: KNN, SVM, FFNN
• Authentication metrics: FAR, FRR, EER
• Automated visualizations and analysis

Quick Start

Requirements: MATLAB R2020b+, Statistics & Machine Learning Toolbox, Deep Learning Toolbox

Run Pipeline:

% Execute all scripts in order
scripts = {'config', 'segment', 'preprocess_features', 'build_dataset', ...
           'train_models', 'evaluate', 'user_thresholds', ...
           'feature_analysis', 'generate_visualizations'};
for i = 1:length(scripts), run(scripts{i}); end

Project Structure

├── config.m                  # Configuration
├── segment.m                 # Window segmentation
├── preprocess_features.m     # Feature extraction (74 features)
├── build_dataset.m           # Normalization & train/test split
├── train_models.m            # KNN, SVM, FFNN training
├── evaluate.m                # Metrics & confusion matrix
├── user_thresholds.m         # Per-user EER thresholds
├── feature_analysis.m        # PCA, ANOVA analysis
├── generate_visualizations.m # All plots
├── data/                     # 10 users × 2 sessions (FD/MD)
└── results/                  # Outputs (metrics, plots)

Data Format

Input: CSV files with 6 columns (Accel XYZ, Gyro XYZ)
Naming: U<ID>NW_<Session>.csv (e.g., U1NW_FD.csv)

• FD = First Day (training)
• MD = Middle Day (testing)

Models

1. KNN - k-nearest neighbors with cross-validation
2. SVM - RBF kernel with ECOC multi-class
3. FFNN - Neural network [128-64-32] layers

Metrics

• Classification: Accuracy, Precision, Recall, F1-Score
• Authentication: FAR, FRR, EER per user
• Visualizations: PCA, confusion matrix, feature importance

Configuration

Edit config.m:

params.window_len = 128;        % ~4s windows @ 32Hz
params.hop = 32;                % 75% overlap
params.normalize_method = 'minmax';
params.ffnn_layers = [128 64 32];

Output

Results folder contains:

• metrics.mat - All performance metrics
• confusion_ffnn.png - Confusion matrix
• pca_2d.png, pca_3d.png - Feature space visualization
• feature_importance.png - Top features
• accuracy_comparison.png - Model performance

Gait recognition using wearable sensor data | November 2025