Abstract
This paper presents a real-time method for extracting information about the locomotive activity of animals in wildlife videos by detecting and tracking the animals' faces. As an example application, the system is trained on lions. The underlying detection strategy is based on the concepts used in the Viola-Jones detector [1], an algorithm that was originally used for human face detection utilising Haar-like features and AdaBoost classifiers. Smooth and accurate tracking is achieved by integrating the detection algorithm with a low-level feature tracker. A specific coherence model that dynamically estimates the likelihood of the actual presence of an animal based on temporal confidence accumulation is employed to ensure a reliable and temporally continuous detection/tracking capability. The information generated by the tracker can be used to automatically classify and annotate basic locomotive behaviours in wildlife video repositories. © 2006 IEEE