Reliability assessment on human activity recognition

Recognition of activity of daily living (ADL) with ubiquitous sensors has been studied so far, aiming to provide services like automatic life logging, elderly monitoring and energy saving in domestic environments. Although existing studies achieve good accuracy of ADL recognition on average, mis-classification of some activities often occur. In this paper, we try to minimize mis-classification in ADL recognition through reliability assessment of the recognition results obtained by machine learning. Specifically, we propose a novel ADL recognition model which extends the random forest classifier trained by ADL data-set by adding the real time uncertainty propagation of the measured variables to each decision tree providing thus the confidence probability of each output class. This adds to the classifier output a confidence value that holds an important role for many purposes such as decision making, features design to improve the classification rate for some classes, etc. The proposed model classifies the input data samples into activity classes with high confidence probability (e.g., more than 50% confidence) and an unclassifiable class, where higher confidence probability leads to the higher recognition accuracy but higher ratio of unclassifiable samples. Through experiments, we confirmed that the proposed model achieve 75% accuracy with less than 30% unclassifiable samples and 95% accuracy with 50% unclassifiable samples.