Academics from the University of Oxford’s Robotics Research Group have devised a novel technique for robust, real-time, visual tracking of previously unseen objects from a moving camera.
It can identify and track new objects in any video sequence, real-time or recorded and regardless of rapid and agile object motion, image blur, varying lighting, camera motion or cluttered and changing backgrounds.
Registration compensates for the linear motion of solid objects while segmentation allows for shape changes and perspective changes that occur when the object turns relative to the camera.
Online learning provides continual refinement of the shape of the object itself and the nature of the background.
The system presents distinct advantages over existing ones as it:
requires minimal knowledge about the target’s shape/appearance but does allow for shape or appearance constraints to be implemented ie it can be set to ignore specific objects or specifically focusing on others
can track targets who appearance changes over time, for example a pulsating heart, spreading fingers, a turning ship or a running animal
is capable of tracking multiple targets at once within a single video image sequence
is supported by real-time implementation
can handle large inter-frame displacements ie can cope with low video frame rates or fast moving objects
requires only a low degree of computational overhead and is suitable for portable and mobile applications
Visual tracking of objects has numerous applications in surveillance (either terrestrial or maritime), military purposes and identification of organs in medical imaging applications.
The technique can be used to control pan-tilt-zoom devices to stabilize a target image or for visual control of a device such as a robot to follow a target or for docking.