Human Activity Recognition with Smartphone Sensors & ML

Human Activity Recognition Using Smartphones

Signal Processing & Machine Learning Project

Ofir Ofek, Kfir | Course: Signal Processing | Dataset: UCI HAR

Motivation & Goal

Application: Healthcare, elderly monitoring, fitness tracking

Opportunity: Leveraging existing smartphone motion sensors

Goal: Build a robust system balancing Accuracy vs. Computational Cost

UCI HAR Dataset Overview

30 Volunteers (Ages 19-48) • 50 Hz Sampling

Walking, Walking Up, Walking Down, Sitting, Standing, Laying

Signal Characteristics

Accelerometor & Gyroscope (3-axial)

Windowing: 2.56 sec segments (128 samples)

50% Overlap for feature continuity

Preprocessing Pipeline

1. Noise Removal (Median Filter)

2. Butterworth Low-Pass Filter (20Hz)

3. Gravity Separation (0.3Hz Cutoff)

4. Feature Extraction

Feature Extraction

561 Features Per Window

Time Domain: Mean, STD, Energy Frequency Domain: FFT, Spectral Energy

Core Research Questions

1. Which classical ML algorithm performs best?

2. Can we achieve >85% accuracy robustly?

3. Trade-off: How does PCA affect Accuracy vs. Speed?

Models Evaluated

SVM (Linear) - High dimensionality handling

K-Nearest Neighbors - Instance-based

Decision Tree - Interpretable logic

PCA: Dimensionality Reduction

• Original: 561 Features (High cost) • 100 Components: 95% Variance • 3 Components: 75% Variance

Results: Baseline (No PCA)

Insight: Linear SVM handles high-dimensional vectors (561 features) most effectively.

Efficiency Trade-off: PCA Impact

CRITICAL: 100 features offer a 57% speed boost with minimal accuracy loss.

Confusion Matrix Analysis

✔ BEST: Laying (~100% Accuracy) - Distinct orientation.

⚠ CONFUSED: Sitting vs. Standing - Similar static posture.

⚠ CONFUSED: Walking vs. Walking Upstairs - Similar dynamic patterns.

Final Conclusions

1. SVM (Linear) is the optimal classical model (88.4%).

2. PCA is viable for real-time systems (minor accuracy trade-off).

3. Static activities are easier to classify than dynamic ones.

Limitations & Future Work

Limitations

• Reliance on hand-crafted features • Offline processing (no real-time stream) • No transition states modeled

Future Work

• Deep Learning (CNN/LSTM) for auto-features • Real-time mobile deployment • User-independent generalization

Thank You

Accurate HAR is achievable with classical ML & Smart Processing.

Questions?