Highly accurate recognition of human postures and activities through classification with rejection

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Human Activity Recognition

Background:

• Accurate posture and activity monitoring is crucial for clinical and research applications.
• Existing methods often struggle to achieve the high accuracy (above 99%) required in some demanding scenarios.
• Wearable sensors offer a promising avenue for continuous and unobtrusive monitoring.

Purpose of the Study:

• To develop and evaluate a method for highly accurate posture and activity recognition using wearable shoe sensors.
• To improve classification accuracy beyond 99% by incorporating a rejection mechanism.
• To compare the performance of different machine learning models and data processing techniques.

Main Methods:

• Utilized data from pressure and acceleration sensors embedded in a wearable shoe monitor (SmartShoe).
• Implemented classification with rejection using Support Vector Machine (SVM) and Multilayer Perceptron (MLP) models.
• Rejected unreliable observations by measuring their distance from the decision boundary against a predefined threshold.

Main Results:

• Achieved a significant increase in classification accuracy from 97.3% ± 2.3% to 99.8% ± 0.1% after applying the rejection method.
• The Multilayer Perceptron (MLP) model, using raw sensor data and rejecting 31.6% of observations, demonstrated superior performance.
• Raw sensor data yielded higher classification accuracy compared to extracted features, and MLP outperformed SVM.

Conclusions:

• Classification with rejection is an effective strategy for enhancing the accuracy of posture and activity recognition from wearable sensors.
• The proposed method, particularly using MLP with raw sensor data, meets the stringent accuracy requirements (>99%) for critical applications.
• This approach is highly beneficial for applications where high recognition accuracy is paramount, even if some data points are not classified.