Machine Observation of the Direction of Human Visual Focus of Attention

摘要: People often look at objects and people with which they are likely to interact. The first step for computer systems adapt the user improve interaction is locate where are, especially location of their faces on image. next track focus attention. For this reason, we interested in techniques estimating tracking gaze people, particular head pose. This thesis proposes a fully automatic approach pose estimation independant person identity using low resolution images acquired unconstrained imaging conditions. developed method demonstrated evaluated densly sampled face image database. We propose new coarse-to-fine that uses both global local appearance estimate orientation. fast, easy implement, robust partial occlusion, no heuristiques can be adapted other deformable objects. Face region normalized size slant by tracker. resulting imagettes projected onto linear auto-associative memory learned Widrow-Hoff rule. Linear memories require very few parameters offer advantage cells hidden layers have defined class prototypes saved recovered all kinds applications. A coarse orientation known unknown subjects obtained searching best prototype matches current search salient facial features relevant each Feature points locally described Gaussian receptive fields intrinsic scale. These descriptors interesting properties less expensive than Gabor wavelets. Salient regions found motivate construction model graph Each node displaced localy according its saliency deliver among neighbors obtains match. associated grid selected as does not use any heuristics, manual annotation or prior knowledge configuration