Human moving behavior estimation from 3-axis accelerometer signal by particle filter with Self-Organizing Map based likelihood

Human moving behaviors such as walking, running, standing, sitting, taking stairs, are to be estimated from 3-axis acceleration signal of smart phone's sensor. For this purpose, state estimation via particle filter with Self-Organizing Map (SOM) based likelihood has been proposed. Input signal is the length of 3 dimensional vector of 3-axis acceleration followed by Gabor Wavelet transform to obtain frequency spectrum. SOM converts the frequency spectrum to a pattern map in two dimension. Then, matching will be conducted for the pattern map with template maps that correspond to the human moving behaviors prepared beforehand the matching. Resulting matching scores will be used as a likelihood in state estimation by particle filter [1]. State space is discrete consisting of human moving behavior with state transition graph, which is equivalently represented in a state transition matrix, being a system model in a framework of state space modeling. The system model restricts possible or likely transition among different behaviors based on the transition matrix. Particle filter algorithm estimates probability over the discrete state space by simulating motion of many samples in the state space. Experiment shows performance of the proposed method for collected data in real scene by wearing a smart phone. Simple input signal and algorithm are advantageous in the proposed method that allows us easy implementation of the method for commercial products.