OBJECT RECOGNITION USING SHAPE AND BEHAVIORAL FEATURES by Chetan Bhole

Object Recognition is the domain of computer vision that deals with the classification of objects in two or three dimensions as instances of predetermined object classes, and is useful for image analysis and understanding. One of the most powerful properties used for recognition of objects is the shape of the object. Other features traditionally used include color, texture, moments and other attributes visible from single static images. The difficulty in obtaining error-free classification is that image data almost always has noise, is cluttered with many different objects and the objects may be occluded or hidden so only a fraction of the object is visible. That is why it is useful to consider additional information which is not available from single static images. In this thesis we define and employ behavioral features along with shape features to characterize objects. We model the dynamic behavior of each object class by its common pattern of movement to help disambiguate different objects that may look similar but belong to different object classes in light of their distinct behavioral features. We show that for a simulated environment problem where we create a database of a mix of objects with small and large variations in shapes, sizes and behavior models, the error rate of the system that uses shape alone decreases from a top-choice error rate of 7-30% to a top-choice error rate of 2-3% using shape and behavioral features, with performance gain with the addition of behavioral features depending on the noise levels in cases where the shapes are similar but objects have different behaviors.