Audio-coupled video content understanding of unconstrained video sequences

Unconstrained video understanding is a difficult task. The main aim of this thesis is to recognise the nature of objects, activities and environment in a given video clip using both audio and video information. Traditionally, audio and video information has not been applied together for solving such complex task, and for the first time we propose, develop, implement and test a new framework of multi-modal (audio and video) data analysis for context understanding and labelling of unconstrained videos. The framework relies on feature selection techniques and introduces a novel algorithm (PCFS) that is faster than the well-established SFFS algorithm. We use the framework for studying the benefits of combining audio and video information in a number of different problems. We begin by developing two independent content recognition modules. The first one is based on image sequence analysis alone, and uses a range of colour, shape, texture and statistical features from image regions with a trained classifier to recognise the identity of objects, activities and environment present. The second module uses audio information only, and recognises activities and environment. Both of these approaches are preceded by detailed pre-processing to ensure that correct video segments containing both audio and video content are present, and that the developed system can be made robust to changes in camera movement, illumination, random object behaviour etc. For both audio and video analysis, we use a hierarchical approach of multi-stage classification such that difficult classification tasks can be decomposed into simpler and smaller tasks. When combining both modalities, we compare fusion techniques at different levels of integration and propose a novel algorithm that combines advantages of both feature and decision-level fusion. The analysis is evaluated on a large amount of test data comprising unconstrained videos collected for this work. We finally, propose a decision correction algorithm which shows that further steps towards combining multi-modal classification information effectively with semantic knowledge generates the best possible results.