An efficient representation of two-dimensional visual objects is specified by an emerging audiovisual compression standard known as MPEG-4. It incorporates the advantages of segmentation-based video compression (whereby objects are encoded independently, facilitating content-based functionalities), and also the advantages of more traditional block-based approaches (such as low delay and compression efficiency). What is not specified, however, is the method of extracting semantic objects from a scene corresponding to a video segmentation task. An accurate, robust and flexible solution to this is essential to enable the future multimedia applications possible with MPEG-4. Two categories of video segmentation approaches can be identified: supervised and unsupervised. A representative set of unsupervised approaches is discussed. These approaches are found to be suitable for real-time MPEG-4 applications. However, they are not suitable for off-line applications which require very accurate segmentations of entire semantic objects. This is because an automatic segmentation process cannot solve the ill-posed problem of extracting semantic meaning from a scene. Supervised segmentation incorporates user interaction so that semantic objects in a scene can be defined. A representative set of supervised approaches with greater or lesser degrees of interaction is discussed. Three new approaches to the problem, each more sophisticated than the last, are presented by the author. The most sophisticated is an object-based approach in which an automatic segmentation and tracking algorithm is used to perform a segmentation of a scene in terms of the semantic objects defined by the user. The approach relies on maximum likelihood estimation of the parameters of mixtures of multimodal multivariate probability distribution functions. The approach is an enhanced and modified version of an existing approach yielding more sophisticated object modelling. The segmentation results obtained are comparable to those of existing approaches and in many cases better. It is concluded that the author’s approach is ideal as a content extraction tool for future off-line MPEG-4 applications.