Visual Tracking: From An Individual To Groups Of Animals

This thesis is concerned with the development and application of visual tracking techniques to the domain of animal monitoring. The development and evaluation of a system which uses image analysis to control the robotic placement of a sensor on the back of a feeding pig is presented first. This single-target monitoring application is then followed by the evaluation of suitable techniques for tracking groups of animals, of which the most suitable existing technique is found to be a Markov chain Monte Carlo particle filtering algorithm with a Markov random field motion prior (MCMC MRF, Khan et al. 2004). Finally, a new tracking technique is developed which uses social motion information present in groups of social targets to guide the tracking. This is used in the new Motion Parameter Sharing (MPS) algorithm. MPS is designed to improve the tracking of groups of targets with coordinated motion by incorporating motion information from targets that have been moving in a similar way. Situations where coordinated motion information should improve tracking include animal flocking, people moving as a group or any situation where some targets are moving in a correlated fashion. This new method is tested on a variety of real and artificial data sequences, and its performance compared to that of the MCMC MRF algorithm. The new MPS algorithm is found to outperform the MCMC MRF algorithm during a number of different types of sequences (including during occlusion events and noisy sequences) where correlated motion is present between targets. This improvement is apparent both in the accuracy of target location and robustness of tracking, the latter of which is greatly improved