Human Activity Recognition with LSTM

Overview

This project focuses on Human Activity Recognition (HAR) using LSTM-based neural networks. The goal is to classify different human activities based on motion sensor data.

Dataset Used

The model is trained on the UCI HAR Dataset, a widely used benchmark dataset for human activity recognition. It contains data collected from accelerometers and gyroscopes of smartphones while subjects performed daily activities.

Model Performance

Classification Report

Below are the precision, recall, and F1-score for each activity class:

              precision    recall  f1-score   support

     Class 0       0.92      0.98      0.95       496
     Class 1       0.95      0.91      0.93       471
     Class 2       0.98      0.95      0.96       420
     Class 3       0.92      0.94      0.93       491
     Class 4       0.94      0.93      0.94       532
     Class 5       1.00      0.99      1.00       537

    accuracy                           0.95      2947
   macro avg       0.95      0.95      0.95      2947
weighted avg       0.95      0.95      0.95      2947

Confusion Matrix

The confusion matrix below visualizes the model's performance in classifying different activities:

Next Steps

Improve the model with GRU & CNN architectures.
Expand testing with real-world sensor data.
Fine-tune hyperparameters for better generalization.