Ablation Analysis to Select Wearable Sensors for Classifying Standing, Walking, and Running

Area of Science:

• Biomechanics and Human Movement Analysis
• Machine Learning Applications
• Wearable Sensor Technology

Background:

• Human activity recognition (HAR) commonly employs wearable sensors and machine learning to identify subject actions.
• Support vector machines (SVM) are frequently used for classification tasks in HAR.

Purpose of the Study:

• To evaluate the accuracy of an SVM for recognizing walking and running activities using principal components from wearable sensor data.
• To perform an ablation analysis to determine the optimal sensor subset for classification.
• To compare principal components across trials to assess trial similarity.

Main Methods:

• Five subjects performed standing, walking, running, and sprinting on a treadmill.
• Data were collected using surface electromyography (sEMG) sensors, inertial measurement units (IMUs), and force plates.
• An SVM was trained on principal components derived from the sensor data, with an ablation analysis conducted.

Main Results:

• Over 90% classification accuracy was achieved for stand, walk, and run/sprint using the first three principal components with all sensors.
• Sensors placed on the lower leg yielded higher accuracies than those on the upper leg.
• Ablating force plates resulted in a minor accuracy decrease, potentially not operationally relevant.
• Excluding sEMG and relying solely on accelerometers reduced SVM accuracy.

Conclusions:

• High classification accuracy for basic human activities is achievable with SVMs and principal components from wearable sensors.
• Sensor placement (lower leg) and modality (inclusion of sEMG) are critical factors for optimizing HAR accuracy.
• The findings suggest that a combination of lower leg sEMG and IMU data provides robust performance for distinguishing gait types.