Contour Tracking

Abstract Object tracking is a fundamental problem in computer vision. It is generally required as a preprocessing step that is used to perform motion‐based object recognition, surveillance, video indexing, autonomous vehicle navigation, and human–computer interaction. In its simplest form, the goal of an object tracker is to generate trajectories of the objects from a sequence of frames. The trajectory can be generated by tracking a single point, such as the object centroid, and by consistently labeling this point in time. Depending on the domain, the object trajectory alone may not be adequate to perform additional tasks, such as recognizing the actions performed by the tracked object. In such scenarios, the object tracker may be required to provide object‐centric information, such as the object shape. Acquiring object‐centric information requires complete object representations. From among various object representations, the object contour tightly encloses the object and is the most suitable representation for tracking the shape of a rigid or nonrigid object. Tracking object contours is achieved by evolving an initial contour from the previous object position to its new position in the current frame. This evolution process is governed by iteratively minimizing a cost function. The cost of the contour at a particular iteration is computed based on the appearance of the object and on the background regions. In addition to the object appearance, a common practice is to impose constraints on the solution by including additional terms in the cost function. These additional terms are related to the smoothness of the resulting curve or the prior shape information of the tracked object.