Visual Animal BiometricsTilo Burghardt, Visual Animal Biometrics. PhD thesis. University of Bristol. June 2008. PDF, 16591 Kbytes. External information
The field of computer vision has repeatedly been recognised as an intellectual frontier whose boundaries of applicability are yet to be stipulated. This thesis explores one novel application: visual animal biometrics.
The work demonstrates that vision can achieve an automatic identification of animals filmed in their natural habitat. The thesis proposes and evaluates~algorithms for detecting a species and - in the case that the animals carry Turing~patterns - for recognising individuals in visual material comprising various poses, changing lighting and clutter. Lions, plains zebras and African penguins serve as sample species to showcase the different capabilities and limitations of the approach.
First, an algorithmic framework for species recognition is discussed. In particular, it is shown that the appearance context of a number of reference points on coat patterns contains a species-specific component, which can be utilised for achieving a robust detection of the investigated animal specimens. The proposed model employs boosted point-surround classifiers as local appearance descriptors.
Second, it is illustrated how pose-normalised texture maps of Turing-patterned coat regions can be extracted based on the extracted reference points. The maps are shown to contain individually-specific features which - using an extension of shape contexts - can be represented as deformation-robust sets of histograms. Finally, it is discussed how distance measures can be used for comparing these sets with population databases to retrieve animal identities.
In order to provide a practical proof of concept, a prototype system was tested in a colony of African penguins and in a preliminary study on a photo collection of plains zebras. The results indicate a system performance that allows for disambiguating individuals with a level of confidence sufficient for tasks of population monitoring.
This outcome marks a promising first step towards an automated, truly non-invasive observation of wild animal populations, which would benefit field-based biology and assist the conservation of species in decline.