A hierarchical model for body height estimation in images

Abstract 

In forensic practice, validation experiments performed on known items or persons are used to make predictions on unknown ones. An example of this is body height estimation in digital images. Using a hierarchical statistical model in this case is quite natural as it allows outcomes of the experiment to depend on random effects for test persons and on fixed effects for operators performing the measurements. In the paper, a hierarchical model is described and implemented in WinBUGS to obtain Bayesian credible intervals for perpetrator heights in a case study involving four perpetrators. Comparing the estimated credible intervals of the Bayesian inference to frequentist confidence intervals proposed in the literature, the results that emerge are quite similar, Bayesian intervals being slightly wider. The hierarchical model takes into account the variation within the individual measurements which is ignored by models using observed means over operators. The approach described is applicable for situations in which on the basis of (repeated) measurements on known objects, a prediction is required on a questioned object under the same circumstances. Another example of this is estimating the speed of a vehicle on video footage on the basis of a validation experiment.

Keywords: Bayesian inference, Body height estimation, Confidence intervals, Credible intervals