Automated 3D Reconstruction of Moving Rigid Specimen from RGB-D Video Input

Abstract

To assist morphometric and behavioural analysis of live animals, we present an automatic process for generating a 3D volumetric representation of a dynamic rigid specimen from RGB-D video. This process uses multi-feature extraction and automatic labelling through a novel distance matrix structure and rigid transformation validation to reduce maximal clique calculations. An adapted SiftFu implementation then incorporates the resulting rigid transformations to perform the final volumetric reconstruction. Validation occurs using an RGB-D sequence from a living leopard tortoise resulting in a smoothly merged and textured volume from the original RGB-D frames.