The recognition of objects and classes of objects is of importance in the field of computer vision due to its applicability in areas such as video surveillance, medical imaging and retrieval of images and videos from large databases on the Internet. Effective recognition of object classes is still a challenge in vision; hence, there is much interest to improve the rate of recognition in order to keep up with the rising demands of the fields where these techniques are being applied. This thesis investigates the recognition of activities and expressions in video sequences using a new descriptor called the spatiotemporal shape context. The shape context is a well-known algorithm that describes the shape of an object based upon the mutual distribution of points in the contour of the object; however, it falls short when the distinctive property of an object is not just its shape but also its movement across frames in a video sequence. Since actions and expressions tend to have a motion component that enhances the capability of distinguishing them, the shape based information from the shape context proves insufficient.

This thesis proposes new 3D and 4D spatiotemporal shape context descriptors that incorporate into the original shape context changes in motion across frames. Results of classification of actions and expressions demonstrate that the spatiotemporal shape context is better than the original shape context at enhancing recognition of classes in the activity and expression domains.