Unsupervised and semi-supervised methods for human action analysis

While human action recognition is a very well studied topic, semi-supervised and unsupervised tasks such as human action retrieval and human action clustering have received relatively little attention. These topics are important to study, as they require far less or no annotated training data, making it more feasible to apply these methods to real-world data, where neatly annotated data are far too rare and costly to obtain. In this thesis, several projects have been undertaken, focused on performing semi-supervised and unsupervised tasks on human actions, with potential for application to more complex systems. The first topic for study is human action retrieval. Various methods for action representation, ranking and relevance feedback are implemented, and compared to one another. The result is a highly accurate human action retrieval system, outperforming the state-of-the-art. This initial investigation is extended with the exploration of human action localisation. Two approaches to this problem are considered. First, a novel, efficient algorithm is introduced for performing temporally unconstrained retrieval and localisation of multimedia human action videos. This algorithm runs several orders of magnitude better than the best contemporary work on several action datasets, while maintaining practical accuracy. Then, a novel algorithm for performing unsupervised temporal localisation of discrete human motions is designed, based on the first two principal components of optical flow. A full human action recognition system is designed around this algorithm to provide an experimental validation of this concept. Experiments show state-of-the-art performance on two popular human action datasets.