Automated visual tracking and social behaviour analysis with multiple mice

Neurodegenerative diseases are characterised by motor deficiencies. For many of them, there is no successful neuroprotective or neuroregenerative therapy clinically available. In order to address this problem, the development of valid animal models for motor disorders has become an active field in preclinical research. Behaviour analysis of laboratory animals is a useful tool to assess therapeutic efficacy. The entire process consists of animal tracking

and behaviour categorisation. Despite efforts made within the research community, there is no system which can perform automated recognition of complex animal behaviours and interactions. In this thesis, we propose to develop a fully automated and trainable computer vision system to track and analyse complex mouse behaviours and interactions using video data recorded by calibrated cameras. To achieve this goal, we firstly propose a novel method based on Baysian-inference Inter Linear Program to continuously track several mice and individual parts without requiring any specific tagging. For automated recognition of singleview mouse behaviours, we present a Hidden Markov Model (HMM) based framework, where the emission probabilities of the HMM are learned by an efficient hybrid architecture including a combination of Segment Fisher Vector and Segment Aggregate Network. Finally

we further extend the system of single-view mouse behaviour recognition for multi-view mouse behaviour recognition based on a deep probabilistic graphical model which jointly describes the unique dynamics from each view, extracts the common pattern across views and represents the correlations between the action labels of the neighbourhoods. We have evaluated the proposed methods by conducting extensive experiments on public and our

datasets. Experimental results show the effectiveness of our methods for tracking, behaviour recognition and social behaviour analysis.