Dynamic models for human activity analysis

Automatic human activity analysis from videos is a very important area of research in computer vision. Of particular interest are the development of a) rich representations for human motion across several domains, 2) algorithms for tasks such as action recognition and tracking, and c) computationally efficient strategies for performing human activity analysis in very large data sets. In this thesis we will address these challenges and propose features and methods for human activity analysis that are very general and can be applied across several domains. The common thread underlying all the methods is the need to explicitly model the temporal dynamics of human motion.;We will first propose an optical-flow based time-series features to model human motion in a scene. We will model the temporal evolution of these features using dynamical systems with stochastic inputs and develop methods for comparing these dynamical systems for the purpose of human activity recognition. We will then address the issue of human tracking by proposing action-specific dynamic templates. The tracking problem will be posed as a joint optimization problem over the location of the person in the scene as well as the internal state of the dynamical system driving the particular activity. Finally, we will propose very fast approximate-nearest neighbor based methods on the space of dynamical systems for analyzing human motion and show that we can perform human activity recognition very efficiently albeit at a cost of slightly decreased accuracy.;Our experimental analysis will show that using dynamical-systems based generative models for human activity perform very well in the above-mentioned tasks.