Signatures of Walking Humans from Passive and Active Acoustic Data using Time-Varying Vector Autoregressions

A sensor fusion framework for characterizing the signature of walking targets using data collected from passive acoustic and active ultrasound sensors is investigated. We compute local estimates of the acoustic energy of the footsteps and the velocity of the torso and limbs. A time-varying vector autoregression (TV-VAR) is used to model the evolution of these signals, and captures the physical correlations between them, creating a natural data fusion across different sensor modalities. The signature is defined as a subset of the parameters from the TV-VAR model, and the quality of this feature set is evaluated using a support vector machine framework to classify multiple test subjects for both detection and discrimination applications.