Loading patent…
AssigneeIsis Innovation Limited
InventorsCharles Colin Bibby, Ian David Reid
Estimating the term…
A method is discussed of tracking objects in a series of n-D images (102) that have objects (106, 108) appearing in a background (110, 112), that method comprises using a probabilistic model of the appearance of the objects and of the appearance of the background in the images, and using an evaluation of whether particular pixels in the images (102) are a part of an object (106, 108) or a part of the background (110, 112), that evaluation comprising determining the posterior model probabilities that a particular pixel (x) or group of pixels belongs to an object or to the background, and further comprising marginalising over these object/background membership probabilities to yield a function of the pose parameters of the objects, where at least the object/background membership is adjudged to be a nuisance parameter and marginalised out.
A method of tracking objects in a series of n-D images that have objects appearing in a background comprising using a probabilistic model of the appearance of the objects and of the appearance of the background in the images, and using an evaluation of whether particular pixels in the images are a part of an object or a part of the background, that evaluation comprising determining the posterior model probabilities that a particular pixel or group of pixels belongs to an object or to the background, and further comprising marginalising over these object/background membership probabilities to yield a function of the pose parameters of the objects, where at least the object/background membership is adjudged to be a nuisance parameter and marginalised out.
A method of segmenting objects in a series of n-D images that have objects appearing in a background comprising using a probabilistic model of the appearance of the objects and of the appearance of the background in the images, and using an evaluation of whether particular pixels in the images are a part of an object or a part of the background, that evaluation comprising determining the posterior model probabilities that a particular pixel or group of pixels belongs to an object or to the background, and further comprising marginalising over these object/background membership probabilities to yield a function of the shape of the objects, where at least the object/background membership is adjudged to be a nuisance parameter and marginalised out.
A method of tracking and segmenting objects in a series of n-D images that have objects appearing in a background comprising using a probabilistic model of the appearance of the objects and of the appearance of the background in the images, and using an evaluation of whether particular pixels in the images are a part of an object or a part of the background, that evaluation comprising determining the posterior model probabilities that a particular pixel or group of pixels belongs to an object or to the background, and further comprising marginalising over these object/background membership probabilities to yield a function of the shape and pose parameters of the objects, where at least the object/background membership is adjudged to be a nuisance parameter and marginalised out.
A method according to any preceding claim wherein a probabilistic appearance model, for example a non-parametric distribution, for image values is used.
A method according to any preceding claim comprising determining the pose of the object in a first image and determining the pose of the object in a second image.
A method according to any preceding claim wherein a level-set of an embedding function is used to represent the object shape, and optionally or preferably wherein a prior condition is imposed on the embedding function to constrain it to be an approximate signed distance function.
An image tracking system comprising:—
An image segmenting system comprising:—
A method according to any preceding method claim further comprising updating the appearance models of both foreground and background.
A method for real-time tracking of a non-rigid foreground object in a sequence of images comprising:
The method of any preceding method claim wherein an object location is initialised by a user defining a region, optionally a closed loop such as a ring or an elliptical region, of foreground to define the appearance model if the object and initial object location, the background being acquired by considering a region outside the user-defined region and wherein steps (iv) and (v) are iterated to extract the shape which best separates background from foreground as determined by maximising P(p,Φ|x,y).
The method of any preceding method claim wherein the object location is initialised by an automatic detection/recognition module (e.g. a face detector)
The method of any preceding method claim wherein the object location in a subsequent image frame is predicted based on its previous location and a model of its expected motion.
The method of any preceding method claim wherein a or the optimisation step for the pose p is achieved using Newton-Raphson technique.
The method of any preceding method claim wherein a or the optimisation step for the shape Φ is achieved using calculus of variations.
The method of any preceding method claim wherein only those pixels are considered that are in a narrow band around the predicted boundary contour separating object from background.
The method of any preceding method claim wherein pose parameters are adjusted using closed-loop control to keep the object centred in its original coordinate frame by accounting for drift caused by shape variations.
The method of any preceding method claim wherein the appearance models are updated via a weighted running average of the previous model and the current appearance.
The method of any preceding method claim further comprising using the best pose parameters p to develop a control signal to control the pose parameters of a robotic device.
The method of any preceding method claim where the best pose parameters p are used to provide a control signal to a virtual world or augmented reality sub-system.
The method of any preceding method claim where the best pose parameters p are used to extract a stabilised view of the foreground object.
The method of preceding method claim in which the images are 3-D, such as MRI or CT, and wherein the probabilistically determined boundary between background and foreground is a surface.
The method of any preceding method claim in which the images are 3-D, such as MRI or CT, and the boundary delineates an interior and an exterior of some object of interest.
The method of any preceding method claim in which the probabilistic appearance model is given by:
The method of any preceding method claim having a model of the appearance of the background and/or object(s) wherein the model of the appearance of the object(s) and/or the model of the appearance of the background can evolve over time.
A visual tracking system for tracking a foreground object in a series of images, that marginalises over a foreground/background segmentation using the posterior probabilities of foreground/background membership of image pixels, and which optimises pose parameters of the object.
A computer program encoded on a computer-readable data carrier which when run on a computer processor performs the method of any of the preceding method claims.
Dependencies are inferred from the claim text.