Transfer Learning Based Recognition of Human Activity Using Wearable Data

This work presents an advanced Human Activity Recognition (HAR) model using transfer learning with an LSTM architecture, leveraging accelerometer data for enhanced prediction accuracy. Utilizing a 200 time-step sequence input, our model is initially trained on the UCI HAR dataset, achieving a strong baseline accuracy of 97%. Further, fine-tuning the WISDM dataset enables the model to adapt to real-time scenarios and achieve 95% accuracy on the test set. The model’s high performance highlights its robustness in accurately identifying between various activities, including walking, jogging, and climbing stairs. Integrating transfer learning with an LSTM-based sequence model demonstrates significant promise for real-time, on-device HAR applications, such as health monitoring and fitness tracking. Our results suggest a reliable and cost-effective method for activity recognition, offering a privacy-preserving alternative to vision-based systems while maintaining high accuracy across diverse activity classes.