Human activity recognition from inertial sensor time-series using batch normalized deep LSTM recurrent networks.

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Machine learning for human activity recognition (HAR) is effective but often requires explicit feature extraction.
• Current methods using body-worn inertial sensors frequently overlook temporal correlations in sequential data.
• This limits the ability to fully capture the dynamics of human movement patterns.

Purpose of the Study:

• To develop a deep recurrent neural network for HAR that processes raw sensor data directly.
• To classify six daily life activities using accelerometer and gyroscope data.
• To evaluate the efficacy of Long Short-Term Memory (LSTM) networks and batch normalization in this domain.

Main Methods:

• Implementation of a Long Short-Term Memory (LSTM) deep recurrent neural network.
• Processing of featureless raw input signals from accelerometer and gyroscope sensors.
• Application of batch normalization to optimize training efficiency.

Main Results:

• The LSTM model achieved 92% average accuracy in classifying six daily life activities.
• The network successfully processed featureless raw input signals, eliminating the need for manual feature engineering.
• Batch normalization reduced the number of training epochs required by approximately four times.

Conclusions:

• LSTM networks offer a powerful approach for human activity recognition directly from raw sensor data.
• The proposed method effectively captures temporal correlations, leading to high classification accuracy.
• Batch normalization significantly enhances the training efficiency of HAR models.