PoseNet: Human Activity Detector

A machine learning pipeline for classifying human physical activities from raw accelerometer and gyroscope sensor signals, achieving 97.2% accuracy using an LSTM network.

Overview

Built on the UCI Human Activity Recognition (HAR) dataset (~10,299 samples, 561 features), this project applies PCA for dimensionality reduction and benchmarks three model architectures — Random Forest, 1D-CNN, and LSTM — across 6 activity classes.

Features

• Dataset: UCI HAR Dataset from 30 subjects performing 6 activities (walking, sitting, standing, etc.)
• Dimensionality reduction: PCA reduces 561 features to 50 components, retaining 98% variance
• Multi-model benchmarking: Random Forest, 1D-CNN, and LSTM trained and evaluated side-by-side
• Cross-validation: 5-fold stratified cross-validation on an 80/20 train-test split
• Training speedup: PCA reduces training time by ~40%

Results

Model	Accuracy	F1-Score
Random Forest	—	—
1D-CNN	—	—
LSTM	97.2%	0.97

LSTM consistently outperformed other architectures across all 6 activity classes.

Activity Classes

1. Walking
2. Walking Upstairs
3. Walking Downstairs
4. Sitting
5. Standing
6. Laying

Tech Stack

• Python
• TensorFlow / Keras — CNN and LSTM models
• Scikit-learn — Random Forest, PCA, cross-validation
• Pandas — data loading and preprocessing
• NumPy — feature matrix operations

How It Works

1. Raw accelerometer and gyroscope signals are loaded from the UCI HAR dataset
2. PCA is applied to compress 561 features into 50 principal components (98% variance retained)
3. Data is split into 80% train / 20% test with stratification across activity classes
4. Three models are trained: Random Forest (on PCA features), 1D-CNN, and LSTM
5. Models are evaluated using accuracy, F1-score, and 5-fold cross-validation

Architecture (LSTM)

Input (time-step sequence × 50 PCA features)
       ↓
LSTM Layer 1
       ↓
LSTM Layer 2
       ↓
Dense → Softmax (6 activity classes)

Usage

git clone https://github.com/<your-username>/posenet-activity-detector
cd posenet-activity-detector
pip install -r requirements.txt

# Preprocess and train
python preprocess.py
python train.py --model lstm

# Evaluate
python evaluate.py --model lstm

Dataset

UCI Human Activity Recognition Dataset — 30 subjects, 6 activity classes, signals sampled at 50Hz. Available at: https://archive.ics.uci.edu/ml/datasets/human+activity+recognition+using+smartphones