Characteristic Classification of Walkers via Underfloor Accelerometer Gait Measurements through Machine Learning

Abstract

The ability to classify occupants in a building has far-reaching applications in security, monitoring human health, and managing energy resources effectively. In this work, gender and weight of walkers are classified via machine learning or pattern recognition techniques. Accelerometers mounted beneath the floor of Virginia Tech's Goodwin Hall measured walkers' gait. These acceleration measurements serve as the inputs to machine learning techniques allowing for classification. For this work, the gait of fifteen individual walkers was recorded via fourteen accelerometers as they, alone, walked down the instrumented hallway, in multiple trials. These machine learning algorithms produce an 88 % accurate model for gender classification. The machine learning algorithms included are Bagged Decision Trees, Boosted Decision Trees, Support Vector Machines (SVMs), and Neural Networks. Data reduction techniques achieve a higher gender classification accuracy of 93 % and classify weight with 64% accuracy. The data reduction techniques are Discrete Empirical Interpolation Method (DEIM), Q-DEIM, and Projection Coefficients. A two-part methodology is proposed to implement the approach completed in this thesis work. The first step validates the algorithm design choices, i.e. using bagged or boosted decision trees for classification. The second step reduces the walking data measured to truncate accelerometers which do not aid in increasing characteristic classification.