Machine learning algorithms can classify outdoor terrain types during running using accelerometry data.

Area of Science:

• Sports Science
• Biomechanical Engineering
• Machine Learning

Background:

• Running is a popular exercise with health benefits.
• Running surface characteristics can impact biomechanical loading and injury risk.
• Wearable motion sensors and machine learning offer potential for automatic running surface identification.

Purpose of the Study:

• To evaluate the accuracy of machine learning algorithms in identifying running surface types using 3D accelerometer data.
• To determine if the number of sensors (single vs. two) affects model accuracy for surface identification.

Main Methods:

• Twenty-nine healthy adults ran on concrete, synthetic, and woodchip surfaces.
• Two 3D accelerometers (lower-back and tibia) recorded running data.
• Gradient Boosting (GB) and Convolutional Neural Network (CNN) models were trained and tested using extracted features and raw signals.

Main Results:

• All tested machine learning models achieved over 90% average accuracy in identifying running surfaces.
• The highest accuracies were observed with a two-sensor Gradient Boosting model (97.0%) and a single tibia-sensor Convolutional Neural Network model (96.1%).

Conclusions:

• Machine learning algorithms can reliably distinguish between running surfaces using data from single or dual accelerometer sensors.
• Automatic surface identification can aid in monitoring training load, enhancing performance, and mitigating injury risk for runners.