Cast shadow modelling and detection

Computer vision applications are often confronted by the need to differentiate between objects and their shadows. A number of shadow detection algorithms have been proposed in literature, based on physical, geometrical, and other heuristic techniques. While most of these existing approaches are dependent on the scene environments and object types, the ones that are not, are classified as superior to others conceptually and in terms of accuracy. Despite these efforts, the design of a generic, accurate, simple, and efficient shadow detection algorithm still remains an open problem. In this thesis, based on a physically-derived hypothesis for shadow identification, novel, multi-domain shadow detection algorithms are proposed and tested in the spatial and transform domains. A novel "Affine Shadow Test Hypothesis" has been proposed, derived, and validated across multiple environments. Based on that, several new shadow detection algorithms have been proposed and modelled for short-duration video sequences, where a background frame is available as a reliable reference, and for long duration video sequences, where the use of a dedicated background frame is unreliable. Finally, additional algorithms have been proposed to detect shadows in still images, where the use of a separate background frame is not possible. In this approach, the author shows that the proposed algorithms are capable of detecting cast, and self shadows simultaneously. All proposed algorithms have been modelled, and tested to detect shadows in the spatial (pixel) and transform (frequency) domains and are compared against state-of-art approaches, using popular test and novel videos, covering a wide range of test conditions. It is shown that the proposed algorithms outperform most existing methods and effectively detect different types of shadows under various lighting and environmental conditions.