Highly Accurate Recognition of Human Postures and Activities Through Classification With Rejection

Abstract

Monitoring of postures and activities is used in many clinical and research applications, some of which may require highly reliable posture and activity recognition with desired accuracy well above 99% mark. This paper suggests a method for performing highly accurate recognition of postures and activities from data collected by a wearable shoe monitor (SmartShoe) through classification with rejection. Signals from pressure and acceleration sensors embedded in SmartShoe are used either as raw sensor data or after feature extraction. The Support vector machine (SVM) and multilayer perceptron (MLP) are used to implement classification with rejection. Unreliable observations are rejected by measuring the distance from the decision boundary and eliminating those observations that reside below rejection threshold. The results show a significant improvement (from 97.3% ± 2.3% to 99.8% ± 0.1%) in the classification accuracy after the rejection, using MLP with raw sensor data and rejecting 31.6% of observations. The results also demonstrate that MLP outperformed the SVM, and the classification accuracy based on raw sensor data was higher than the accuracy based on extracted features. The proposed approach will be especially beneficial in applications where high accuracy of recognition is desired while not all observations need to be assigned a class label.