🧠 Human Activity Recognition (Walk vs Run) — Machine Learning & Deep Learning

This project focuses on classifying human activities (walking vs running) using accelerometer and gyroscope sensor data.

It demonstrates an end-to-end data science pipeline — from data cleaning, EDA, and feature engineering to building, evaluating, and saving Machine Learning (ML) and Deep Learning (DL) models.

The final LSTM (Long Short-Term Memory) model achieved nearly 99.9% accuracy, outperforming traditional ML methods and proving its effectiveness in modeling temporal dependencies in sensor data.

---

🔹 Impact Summary

• 🚀 99.9% Accuracy achieved using LSTM sequence modeling
• 🤖 Combined Machine Learning and Deep Learning pipelines for comparison
• 🩺 Applicable to wearable tech, healthcare monitoring, and IoT safety systems
• 💾 Deployment-ready models (.joblib, .h5) stored in output models/

---

🚀 Key Highlights

• 🔍 Conducted comprehensive EDA on multi-axis time-series data to uncover distinct walking and running signal patterns.
• 🧹 Engineered magnitude-based features (accel_mag, gyro_mag), boosting model accuracy by ~4%.
• ⚙️ Designed a unified ML/DL training pipeline comparing Logistic Regression, Random Forest, MLP, and LSTM.
• 📈 Achieved 99.9% test accuracy with the LSTM model by effectively capturing temporal dependencies.
• 🧠 Evaluated using Precision, Recall, F1, ROC-AUC, confirming strong model generalization.
• 💾 Saved trained models in .joblib and .h5 formats for reproducible deployment.
• 🧩 Framework easily extendable to multi-class HAR tasks (sitting, climbing, jumping).
• 🏭 Demonstrated real-world impact in healthcare, wearables, and industrial monitoring.

---

🧩 Problem Statement

In wearable and smart health applications, devices continuously record motion data from sensors like accelerometers and gyroscopes.

The challenge is to automatically differentiate activities (e.g., walking vs running) based on subtle differences in motion patterns.

Goal:
Develop a robust classification system capable of detecting whether a person is walking or running using only raw 3-axis sensor data, while ensuring high generalization across users and devices.

---

📁 Project Structure

├── data
│ └── walkrun.csv
│
├── notebook
│ └── WalkRunClass_LSTM_completed.ipynb
│
├── output models
│ ├── scaler.joblib
│ ├── random_forest.joblib
│ ├── mlp_model.h5
│ └── LSTM_Model.h5
│
├── .gitignore
└── README.md

📊 Data Understanding

The dataset contains timestamped accelerometer and gyroscope readings across three axes (x, y, z) for two activities — walking and running.

Key Observations from EDA:

• Running shows higher acceleration and gyroscope variance.
• Strong correlation between accel_y and gyro_y, representing forward motion.
• The dataset is balanced: ≈50% walk, 50% run.

---

⚙️ Feature Engineering

Feature	Description	Purpose
accel_mag	√(x² + y² + z²)	Total acceleration magnitude
gyro_mag	√(x² + y² + z²)	Total angular velocity magnitude
Normalized features	StandardScaler	To stabilize training and improve convergence

---

🤖 Modeling Approach

Four models were developed and compared:

Model	Type	Framework	Key Strength
Logistic Regression	ML	Scikit-learn	Baseline linear model
Random Forest	ML	Scikit-learn	Handles non-linear interactions
MLP (Neural Network)	DL	TensorFlow/Keras	Dense layers for pattern extraction
LSTM	DL	TensorFlow/Keras	Captures temporal dependencies

---

📊 Model Comparison & Results

Model	Accuracy	Precision	Recall	F1 Score	AUC
Logistic Regression	0.9525	0.9687	0.9354	0.9517	0.9900
Random Forest	0.9884	0.9936	0.9861	0.9884	0.9991
LSTM (Deep Learning)	0.9998	0.9998	0.9995	0.9992	0.9995
MLP (Neural Net)	0.9747	0.9958	0.9676	0.9815	0.9963

✅ Best Model: LSTM — effectively captured sequential temporal dependencies in sensor data, leading to superior accuracy and generalization.

---

🔍 Interpretation & Insights

• LSTM learned temporal motion patterns more effectively than static ML models.
• Random Forest provided explainable feature importance — accel_y & gyro_y dominate.
• Consistently high AUC (>0.99) confirms strong class separation and reliability.

---

🧪 Experimental Learnings

• Feature engineering significantly impacts sensor data performance.
• Deep learning outperforms ML for time-series due to sequence awareness.
• Proper scaling and stratified splits stabilize generalization.
• Random Forest remains a great interpretable baseline.

---

🏭 Business & Research Applications

• 🏃‍♂️ Fitness Trackers: Automatic activity recognition.
• 🏥 Healthcare: Mobility monitoring & rehabilitation tracking.
• 🏭 Industrial Safety: Detecting abnormal or hazardous motions.
• 🧠 Smart Devices: Gesture recognition & IoT context-awareness.

---

🧭 Future Enhancements

• Expand to multi-class activities (sitting, jumping, climbing).
• Integrate CNN-LSTM hybrid for longer time windows.
• Deploy real-time Streamlit dashboard.
• Convert to TensorFlow Lite for mobile inference.
• Serve predictions via FastAPI REST API.

---

🛠️ Tech Stack

Category	Tools / Libraries
Language	Python
ML Frameworks	Scikit-learn, TensorFlow, Keras
Data Handling	Pandas, NumPy
Visualization	Matplotlib, Seaborn
Model Persistence	Joblib, H5
Version Control	Git, GitHub
IDEs	Jupyter Notebook, VS Code

---

🚀 How to Run Locally

Clone the Repository

git clone https://github.com/Bharath-19/Human-Activity-Recognition-Walk-Run.git
cd Human-Activity-Recognition-Walk-Run

👨‍💻 Author

Veera Bharath Chandra Budhi
🎓 Master’s in Autonomy Technologies — Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)
💼 Data Scientist

📫 Contact:

• 📧 [email protected]

💬 Data Scientist skilled in building end-to-end machine learning and deep learning solutions — from data analysis and feature engineering to model deployment and visualization for real-world decision making.

---

⭐ If you found this project helpful, please star the repository!