Moving the Lab into the Mountains: A Pilot Study of Human Activity Recognition in Unstructured Environments

Area of Science:

• Wearable technology
• Biomedical engineering
• Machine learning

Background:

• Human activity recognition (HAR) is crucial for monitoring health and performance.
• Existing HAR methods often rely on controlled lab environments, limiting real-world applicability.
• Mountainous environments present unique challenges due to terrain variability and participant fatigue.

Purpose of the Study:

• To develop and validate a field-based data collection method for HAR in mountains.
• To assess the impact of terrain variations and fatigue on HAR accuracy.
• To train a deep learning model for robust HAR in challenging outdoor conditions.

Main Methods:

• A protocol was created to generate an unsupervised, labelled dataset of mountain activities (running, walking, obstacle climbing).
• Data collection incorporated diverse terrains (slopes, rivers, varied surfaces) and fatigue levels (rested to exhaustion).
• A convolutional neural network (CNN) was trained on this dataset and compared against a lab-based dataset.

Main Results:

• The trail running dataset comprised 3,829,759 samples.
• The CNN achieved high accuracy (0.978) and precision (0.802 for climbing gate) after hyperparameter tuning.
• Field-based HAR accuracy (97.8%) closely matched lab-based performance (97.7%).

Conclusions:

• The study demonstrates the first successful HAR in a mountain environment using a single accelerometer and deep learning.
• A robust protocol was established for creating validated datasets in uncontrolled, real-world conditions.
• The developed deep learning model shows excellent performance across varied terrain and fatigue levels.