Tracking multiple animals in wildlife footage

Abstract

We describe a method for tracking animals in wildlife footage. It uses a CONDENSATION particle filtering frame-work driven by learnt characteristics of specific animals. The key contribution is a periodic model of animal motion based on the relative positions over time of trackable features at significant body points. We also introduce techniques for maintaining a multimodal state density within the particle filter over time to enable consistent tracking of multiple animals. Initial experiments show that the approach has considerable potential.