Person tagging in an image processing system utilizing a statistical model based on both appearance and geometric features

Claim 1

A method of person tagging in an image processing system, the method comprising the steps of: processing a sequence of images to generate a statistical model for each person to be tagged, the statistical model incorporating at least one appearance feature and at least one geometric feature of the tagged person; applying the model to at least one subsequent image in order to perform at least one of a detection operation, a location operation and a tracking operation for the tagged person; and controlling an action of the image processing system based on a result of the at least one operation; wherein the statistical model generated for a given person Ω and image I comprises a likelihood probability function P ⁡ ( I ❘ T , ξ , Ω ) = ∑ pix ∈ I ⁢ P ⁡ ( pix ❘ T , ξ , Ω ) , where r is an index to regions of similar appearance and N is a total number of such regions, r=1, 2, . . . N, is a linear transformation used to capture global motion of the person in the image space and P ⁡ ( pix ❘ T , ξ , Ω ) = max ( P ⁡ ( pix r = 1 , ⁢ … ⁢ , ⁢ N ❘ r , T , ξ , Ω ) ⁢ P ⁡ ( r ❘ ξ , Ω ) ) , where P(pix|r, T, ξ, Ω) is the probability of observing pixel pix assuming that it belongs to an r-th region of the model on a pose ξ, and P(r|ξ, Ω) is the prior probability of the region at that pose.

Claim 2

A method of person tagging in an image processing system, the method comprising the steps of: processing a sequence of images to generate a statistical model for each person to be tagged, the statistical model incorporating at least one appearance feature and at least one geometric feature of the tagged person; applying the model to at least one subsequent image in order to perform at least one of a detection operation, a location operation and a tracking operation for the tagged person; and controlling an action of the image processing system based on a result of the at least one operation; wherein the statistical model generated for a given person Ω and image I comprises a likelihood probability function P ⁡ ( I ❘ T , ξ , Ω ) = ∑ pix ∈ I ⁢ P ⁡ ( pix ❘ T , ξ , Ω ) , where r is an index to regions of similar appearance and N is a total number of such regions, r=1, 2, . . . N, is a linear transformation used to capture global motion of the person in the image space and P ⁡ ( pix ❘ T , ξ , Ω ) = max ( P ⁡ ( pix r = 1 , ⁢ … ⁢ , ⁢ N ❘ r , T , ξ , Ω ) ⁢ P ⁡ ( r ❘ ξ , Ω ) ) , where P(pix|r, T, ξ, Ω) is the probability of observing pixel pix assuming that it belongs to an r-th region of the model on a pose ξ, and P(r|ξ, Ω) is the prior probability of the region at that pose.