Data Pre-processing and Model Selection Strategies for Human Posture Recognition

Accelerometers are widely used in many different physical activity surveillance projects. The goal of this study is to investigate better strategies and algorithms for data preprocessing and model training/testing, based on a single writsworn accelerometer. In the data pre-processing section, 6 sampling rates, five feature selection techniques and three feature scaling methods, as well as 14 sub-feature sets were compared and selected via three classification algorithms (NN, SVM, and Bayes). Then three types of models (data from single subject, part of combined subjects and full combined subjects) were trained and compared based on three testing sets (collected from one training subject, 5 training subjects, and 12 new subjects respectively). Moreover a plurality voting mechanism was applied to adjust the original prediction result during the model testing stage. Finally, an individual model with a plurality voting mechanism were suggested for improving the robustness and reliability of an overall system based on our experimental results.