IMU-Based Movement Trajectory Heatmaps for Human Activity Recognition.

Area of Science:

• Computer Science
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Ubiquitous computing drives human activity recognition (HAR) using inertial sensor data.
• Limited annotated data, particularly in healthcare, hinders HAR model performance.
• Data augmentation techniques are effective in overcoming data scarcity in image classification.

Purpose of the Study:

• To evaluate the effectiveness of transforming inertial sensor data into movement trajectories and 2D heatmap images for HAR.
• To assess the benefits of this approach in scenarios with limited annotated data, such as in healthcare.
• To explore the utility of visual representations for analyzing human motion patterns.

Main Methods:

• Inertial sensor data (acceleration, orientation, angular velocity) were transformed into movement trajectories and 2D heatmap images.
• A convolutional long short-term memory (ConvLSTM) network was employed to classify the generated heatmap images.
• The methodology was evaluated using the Deep Inertial Poser (DIP) dataset.

Main Results:

• State-of-the-art methods remain optimal for large datasets with numerous subjects.
• A performance advantage was observed for smaller datasets, typical in healthcare applications.
• Movement trajectories offered a valuable visual tool for interpreting and analyzing human motion patterns.

Conclusions:

• Transforming inertial sensor data into heatmap images via data augmentation can enhance HAR performance, particularly when training data is scarce.
• This visual representation aids in understanding and analyzing human activities, offering benefits for research in fields like healthcare.
• The ConvLSTM network effectively leverages spatiotemporal information from heatmap images for accurate HAR.