Sensor Input Type and Location Influence Outdoor Running Terrain Classification via Deep Learning Approaches.

Area of Science:

• Biomechanics and Human Movement Analysis
• Machine Learning in Sports Science
• Wearable Sensor Technology

Background:

• Optimizing running performance and preventing injuries requires understanding how different running surfaces affect biomechanics.
• Deep learning models, specifically convolutional neural networks (CNNs), show promise for classifying activities using body-worn sensors.
• No prior research has optimized signal type, sensor location, and model architecture for classifying running surfaces.

Purpose of the Study:

• To identify the optimal combination of signal type, sensor location, and CNN architecture for accurately classifying grass and asphalt running surfaces.
• To evaluate the impact of preprocessing steps and data splitting protocols on classification accuracy.

Main Methods:

• Collected full-body inertial measurement unit (IMU) data from 40 runners on grass and asphalt surfaces.
• Tested various signal types (acceleration, angular velocity), sensor configurations (full body, lower body, pelvis, feet), and CNN architectures.
• Assessed the influence of preprocessing (gait cycle separation, amplitude normalization) and data splitting methods (leave-n-subject-out, subject-dependent).

Main Results:

• Acceleration signals outperformed angular velocity, improving classification by 3.8%.
• The foot sensor configuration achieved the highest accuracy (95.5%) relative to the number of sensors used.
• Separating data into gait cycles and avoiding amplitude normalization significantly boosted accuracy by approximately 28%.

Conclusions:

• Key parameters for developing effective machine learning classifiers for human activity recognition have been identified.
• A running surface classification tool can offer valuable feedback to athletes and coaches for training personalization and injury prevention.
• This technology has the potential to enhance running performance by providing quantitative insights into technique and effort across different terrains.