Non-parametric Bayesian human motion recognition using a single MEMS tri-axial accelerometer.

Area of Science:

• Human-computer interaction
• Wearable sensor technology
• Machine learning for activity recognition

Background:

• Human motion recognition is crucial for applications like healthcare and sports analytics.
• Existing methods often require labeled training data, limiting their applicability in real-world scenarios.
• Unsupervised learning offers a promising alternative for motion recognition without prior data collection.

Purpose of the Study:

• To propose a non-parametric clustering method for recognizing human motions using a single microelectromechanical system (MEMS) accelerometer.
• To enable motion recognition without a priori knowledge of the number of motions or the need for training data.
• To evaluate the proposed method's accuracy against established clustering techniques.

Main Methods:

• Feature extraction from a single MEMS accelerometer.
• Application of the infinite Gaussian mixture model (IGMM) for non-parametric clustering.
• Utilizing a collapsed Gibbs sampler for efficient model inference.
• Comparison with Fuzzy C-Mean (FCM), K-means, and mean-shift algorithms.

Main Results:

• The proposed non-parametric method accurately detects and recognizes unanticipated human motions.
• Significant accuracy improvements were observed compared to parametric (FCM) and unsupervised (K-means, mean-shift) methods.
• The method effectively clusters human motions without requiring predefined motion categories or training datasets.

Conclusions:

• The developed non-parametric clustering approach offers a robust and accurate solution for human motion recognition using MEMS accelerometers.
• This unsupervised technique eliminates the need for labeled data, making it highly practical for diverse applications.
• The IGMM-based method demonstrates superior performance in identifying and classifying human activities in an unknown environment.