IMU-Based Fitness Activity Recognition Using CNNs for Time Series Classification.

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Mobile fitness applications rely on accurate activity tracking via inertial measurement units (IMUs).
• Convolutional neural networks (CNNs) excel at time-series classification but face challenges in fitness activity recognition (FAR) due to data scarcity and activity similarity.
• Human activity recognition (HAR) tasks often utilize traditional machine learning (ML) methods.

Purpose of the Study:

• To evaluate the effectiveness of CNNs for fitness activity recognition (FAR) using IMU data.
• To determine the impact of input data size and sensor count on CNN performance in FAR.
• To compare CNN performance against traditional ML methods for FAR.

Main Methods:

• Adapted three existing CNN architectures and developed a novel Scaling-FCN for FAR.
• Implemented a preprocessing pipeline and collected a running exercise dataset from 20 participants.
• Evaluated four CNNs and three traditional ML methods (including Support Vector Machines - SVMs) on the collected dataset.

Main Results:

• All CNN architectures achieved over 94% test accuracy.
• Traditional ML methods, particularly SVMs, outperformed CNNs in the default scenario (99.00 ± 0.34% accuracy).
• Reducing sensors improved CNN performance, with the Scaling-FCN reaching 99.86 ± 0.11% accuracy using a single foot sensor.

Conclusions:

• CNNs are suitable for fitness activity recognition (FAR) using IMU data.
• Selective sensor reduction can significantly enhance CNN performance in FAR.
• Traditional ML methods remain competitive, especially with ample, favorable input data.