Individual Animal Identification using Visual Biometrics on Deformable Coat Patterns

  • Tilo Burghardt
  • Neill Campbell
Schlagworte: visual biometrics, spatial models, deformable object recognition, real-world application, DDC: 004 (Data processing, computer science, computer systems)

Abstract

In this paper we propose and evaluate an approach to the so far unsolved problem of robust _individual_ identification of patterned animals based on video filmed in widely unconstrained, natural habitats. Experimental results are presented for a prototype system trained on African penguins operating in a real-world animal colony of thousands. The system exploits the individuality of Turing-like camouflage patterns as identity cues since, for a wide range of species, these contain highly unique and compact distributions of phase singularities. The key problem solved in this paper is a distortion robust detection and individual comparison of non-linearly _deforming_ animals. We address the problem using a coarse-to-fine methodology that task-specifically extends and combines vision techniques in a three-stage approach: 1) Using a recently suggested integration of multiple instances of appearance detectors (Viola-Jones) and sparse feature trackers (Lucas-Kanade), a coarse, robust real-time detection of animals in appropriate poses is achieved. 2) An estimation of the 3D-deformed pose is derived by a fast, guided search on a precalculated pose-configuration model, we refer to as _Feature Prediction Tree_, which is learned off-line based on an animated, deformable 3D species-model. The estimate is refined using bundle adjustment posing a polygonal model into the scene. Following, a back-projection of the visible animal surface yields a normalised 2D texture map. 3) An extended variant of _Shape Context_ descriptors (built from filter-extracted phase singularities of characteristic texture areas) are employed as biometric templates. Finally, a distortion-robust identification is achieved by solving associated bipartite graph matching tasks for pairs of these descriptors.
Veröffentlicht
2007-12-31
Rubrik
The 5th International Conference on Computer Vision Systems