Individual Animal Identification using Visual Biometrics on Deformable Coat Patterns

In this paper we propose and evaluate a recognition approach to individual animal identification in patterned species based on video filmed in widely unconstrained, natural habitats. The key issue addressed is a distortion robust detection and comparison of unique, deforming camouflage markings as found in a wide range of species. We propose a coarse-to-fine methodology specifically extending and combining vision techniques in a three-stage approach, that is 1) a rapid, coarse key-view detection based on patch appearance, 2) pose estimation and 3D model fitting using a (pre-computed) dynamic Feature Prediction Tree (FPT)followed by bundle adjustment and 3) texture back-projection, extraction of unique phase singularities and final encoding using an extended variant of Shape Contexts. Distortion-robust animal identification is then achieved by solving associated bipartite graph matching tasks for pairs of templates. Independently producing time-stamped identification data, the system marks a first step towards a partial automation of biological field observations that may permit for a truly non-intrusive behavioural as well as conservational analysis of population dynamics.